Paige.AI, Inc.
[2024] APO 46
•18 November 2024
IP AUSTRALIA
AUSTRALIAN PATENT OFFICE
Paige.AI, Inc. [2024] APO 46
Patent Application: 2020421609
Title:Systems and methods for analyzing electronic images for quality control
Patent Applicant: Paige.AI, Inc.
Delegate:Dr David Carberry
Decision Date: 18 November 2024
Hearing Date: 15 February 2024, by Written Submissions
Catchwords: PATENTS – Examiner Objection hearing – s 45 – whether claims lack a manner of manufacture – artificial intelligence – use of machine learning models – quality control and assurance of pathology images – whether substance lies in an improved method of analysing images – substance lies in an administrative scheme to automate the role of a pathologist – all claims lack a manner of manufacture – no patentable subject matter identified in the specification – application refused
Representation: Patent attorneys for the applicant: Rory Campbell and Ross Clark of Davies Collison Cave
IP AUSTRALIA
AUSTRALIAN PATENT OFFICE
Patent Application: 2020421609
Title:Systems and methods for analyzing electronic images for quality control
Patent Applicant: Paige.AI, Inc.
Date of Decision: 18 November 2024
DECISION
The claimed invention is not for a manner of manufacture. I see no subject matter in the specification which could be made the subject of a valid claim to overcome this finding. I refuse the application.
REASONS FOR DECISION
Background
Paige.AI, Inc (the Applicant) filed US application 62/957517 on 06 January 2020. This was used as the priority document for international application WO2021/141757. AU application 2020421609 (the present application) entered national phase from following this on 26 July 2022. Consequently, it derives an earliest priority date of 06 January 2020.
The Applicant requested the application be examined, and that it be expedited under the Patent Prosecution Highway. They also submitted the Notice of Allowance from the USPTO and amendments to bring the AU application into line with the US application.
Four examination reports followed. The first examination report, dated 25 October 2022, objected to all of the claims on the grounds of inventive step and on a lack of a manner of manufacture. The objection on the basis of inventive step was withdrawn in the second examination report, dated 03 May 2023. However, the manner of manufacture objection was maintained in each of the examination reports. The third and fourth examination reports were issued on 18 August 2023 and 12 October 2023 respectively.
A hearing was requested on 24 October 2023 following the fourth examination report. The Applicant’s written submissions (AWS) were received on 15 February 2024 along with a request to amend the specification under s104. A correction to paragraph [061] of the AWS was received on 15 May 2024 and is also considered as part of this decision.
As the Application was filed after 15 April 2013 it is governed by the Patents Act 1990 (the Act) and Patents Regulations 1991 as amended by the Intellectual Property Laws Amendment (Raising the Bar) Act 2012. The standard of proof that applies in the present case is the balance of probabilities. Under Reg 13.4 I may extend the time to gain acceptance following a decision of the Commissioner. However, I may refuse the application under s49(2) if I find that the specification does not satisfy the criteria for acceptance and is unlikely to do so.[1]
[1] the Act, s 49(1)
I have compared the proposed amendments filed 15 February 2024 with the disclosure in the specification, particularly paragraphs [028]-[029], [034], [058], and am satisfied that the proposed amendments meet the requirements of s102(1) of the Act. While the amendments appear allowable, I have reservations regarding whether the proposed claim amendments and the specification would satisfy s40(2)(a) and s40(3) of the Act. That is, if there is any s40(2)(a) or s40(3) deficiency it is not a result of the amendment. Thus, in this decision I am considering the specification as proposed to be amended on 15 February 2024. Whenever I refer to the specification, it is a reference to the specification as proposed to be amended.
The invention as described
Pathology specimens, such as biopsies, are used to determine whether diseases, such as cancer, are present within the specimen.[2] Specimens are often sectioned, smeared, etc and added onto microscope slides. Stains are usually added to these specimens and these bind to different components within the specimen. Stains may add colour, fluorescence, or other ways to identify the different components.[3] Following these processes, the slides are imaged and the various types of stains are used to assist in identifying whether a specimen contains a type of disease and, if so, what the disease is.[4] The analysis of pathology images is typically performed by trained pathologists or similar professionals. The specification identifies that to provide quality assurance and avoid incorrect diagnoses for patients, a second pathologist reviews particular types of outcomes, including first-time cancer diagnoses, changed diagnoses such as remission, and sometimes merely randomised sampling.[5]
[2] Specification at [002]-[003]
[3] Ibid at [028]
[4] Ibid at [029]
[5] Ibid at [002]-[003]
As indicated by the specification, it is known that quality control (QC) of samples and quality assurance (QA) of diagnoses can be improved, but that it is often impractical, inefficient and costly due to the requirement for two pathologists to review the images. It is in this context that the Applicant aims to provide a computer-based, artificial-intelligence method to supply quality control information to pathologists and to provide quality assurance to supplement specimen evaluation and diagnosis.[6]
[6] Ibid at [002]-[003]
From a technical point of view, the specification is pitched at a high level. Functional language has been employed extensively to define the inputs to particular features, followed by the desired result. Most of the technical detail has been left as an exercise for the person skilled in the art to implement.
10. The specification identifies that a machine learning model may be trained to determine the quality of images, and to output a quality designation of the specimen. This is referred to as a “QC machine learning model”.[7] Any suitable machine learning model may be used, including:[8]
“Convolutional Neural Networks (CNN), CNN with multiple-instance learning or multi-label multiple instance learning, Recurrent Neural Networks (RNN), Long-short term memory RNN (LSTM), Gated Recurrent Unit RNN (GRU), graph convolution networks, or the like or a combination thereof.”
[7] Ibid at [033]-[034]
[8] Ibid at [061]
11. Traditional computer vision analysis techniques may also be utilised in combination with the above machine learning models. The specification states that the choice of machine learning model is largely determined by how much training data is available.[9] It assumes that the selection of an appropriate number of training images, and therefore model, is something which the skilled addressee is fully capable of.
[9] Ibid at [061]
12. The process of training the machine learning model is outlined in Figure 3 (presented below). Training data 302 consists of pathology images, known outcomes, and additional input data related to other factors (e.g. the medical history of the patient, types of stains, pathologist, etc). This training data is input into training algorithm 310, and it supplies outputs called “quality designations”. As the model learns (improves) it is expected that the quality designations will match the known outcomes. Exactly how the machine learns depends on the type of machine learning model utilised. The machine learning algorithm post-training is the QC machine learning model.[10] Following training, when pathology images and additional data are input into the QC machine learning model it is expected that a quality designation will be output. This output is expected to be related to the amount of learning and quality of input data the system was trained with. Example quality designations are described as a specimen approval, a rejection, or a quality score.[11]
[10] Ibid at [062]
[11] Ibid at [064]-[065]
13. A “QA machine learning model” is obtained by following the same general procedure, but by inputting different training data. The output of the QA machine learning model is called a “disease designation”.[12] The outputs of the QC machine learning model and the QA machine learning model may be sent to various users through standard means, including via notifications, visual indicators, and reports.[13]
[12] Ibid at [035]
[13] Ibid at [038]
14. Several paragraphs describe that the computer used to implement the machine learning models may be physically located either on-site or remotely, and that it may connect to other computers via a standard computer network.[14] The computer may physically be any type of computer, ranging from a cloud-based implementation (e.g. server) through to an on-site laptop or handheld mobile device.[15] Input Data 306 may be sourced from any networked computer, and comprise information including “patient-specific information, such as age, medical history, cancer treatment history, family history, past biopsy or cytology information, etc.”[16] Similarly, the pathology images may be shared to any of the computers on the network. Apart from the specific names of the computers (physician servers, hospital servers, clinical trial servers, research lab servers, and laboratory information systems), there is nothing beyond an ordinary networked computer implementation.
[14] Ibid at [039], [041]-[044]
[15] Ibid at [043], [051], [053]
[16] Ibid at [043]
15. Figure 1B (reproduced below) is used as an example of the components and functions which one or both of the QC/QA machine learning models would utilise. Regarding the items within the Figure, data ingestion tool 102 is described in paragraph [048] as:
“The data ingestion tool 102 refers to a process and system for facilitating a transfer of the digital pathology images to the various tools, modules, components, and devices of the machine learning module 100 that are used for characterizing and processing the digital pathology images, according to an exemplary embodiment.”
16. Slide intake tool 103 is described as:[17]
“The slide intake tool 103 refers to a process and system for scanning pathology images and converting them into a digital form, according to an exemplary embodiment. The slides may be scanned with slide scanner 104, and the slide manager 105 may process the images on the slides into digitized pathology images and store the digitized images in storage 106.”
[17] Ibid at [049]
17. QA/QC Tool 101 loads the QC and/or QA machine learning model, obtains the data from the data ingestion tool 102 and the slide intake tool 103, and processes the data.[18] It is then able to output the results to the viewing application tool 108, or send the results to external networked computers.
[18] Ibid at [047]
18. The viewing application tool 108 may provide information which would assist a pathologist to also characterise the specimen. This includes the pathology image, overlays for areas of concern/interest, and information output from the QC/QA machine learning models.[19]
[19] Ibid at [050]
19. Figure 2A provides a flow chart for operating the QC machine learning model. Steps 202, 204, 206, 208 and 210 are self-explanatory. Paragraphs [054]-[061], [065]-[070] link the steps to how the machine learning models are trained and the desired functions of the machine learning modules.
20. Following the determination of a quality designation, it is output from the QC machine learning model and can be saved, or used in some form of notification, such as a report.[20]
[20] Ibid at [065]-[068]
21. Figure 2B provides a flow chart for using the QC and QA machine learning models in sequence. Again, the flowchart is self-explanatory, noting that step 224 effectively incorporates all the steps of Fig 2A.[21]
[21] Ibid at [071]-[080]
22. It is described that the external designation of steps 232 and 234 may be a manual designation provided by a pathologist or it may be based on an artificial digital image.[22]
[22] Ibid at [081]-[083]
23. Later paragraphs describe how the above QC and QA machine learning models may be used in a clinical workflow.[23] The paragraphs describe that it may be performed prior to, in parallel with, or after a pathologist evaluates the specimen. When it has been trained to do so, the QC and QA machine learning models also process non-image data, such as a patient record, prior to the preparation of the pathology specimens.
[23] Ibid at [089]-[097]
24. Figures 5 to 7 (not reproduced here) and the paragraphs [098]-[0117] provide and explain simple flow diagrams where one of the outputs from the QC or QA machine learning models is used to determine further actions. For the QC model, these actions can include saving the file, using the image for a diagnosis, generating notifications/alerts, or recapturing the image. Similarly, the QA machine learning model may output data which is incorporated into reports, save or send the data, and compare against an “external designation” and potentially generate notifications of various types.
25. Finally, paragraphs [0118] – [0122] detail the hardware features of standard computer systems, such as the provision of processors, memory, networking, and storage. Paragraph [0122] explicitly states in part that “[t]he hardware elements, operating systems and programming languages of such equipment are conventional in nature, and it is presumed that those skilled in the art are adequately familiar therewith.”
Claims
26. The specification ends with 18 claims, of which Claims 1, 7, and 18 are independent claims. Independent claim 1 is as follows:
1. A computer-implemented method for performing quality control (QC) and quality assurance (QA) of a pathology specimen, the method comprising:
receiving a digital image corresponding to a target specimen associated with a pathology category, wherein the digital image is an image of human or animal tissue and/or an image algorithmically generated to replicate human or animal tissue;
determining a QC machine learning model, the QC machine learning model being generated by processing a plurality of training images, associated with the pathology category, to predict a quality designation based on one or more artifacts,
wherein the one or more artifacts comprise a missing tissue, a glass crack, a bubble, a blur amount, a missing tissue, a folded tissue, a line, a scratch, dust, pen, or a stain amount;
providing the digital image as an input to the QC machine learning model;
receiving the quality designation for the digital image as an output from the QC machine learning model;
determining a QA machine learning model different than the QC machine learning model, the QA machine learning model being generated by processing a plurality of training images, associated with the pathology category, to predict a disease designation based on one or more biomarkers,
wherein the biomarkers are indicative of one or more an over-expression of a protein, a gene product, an amplification of a specific gene, or a mutation of a specific gene, and wherein the biomarkers are not reliably identifiable using visual analysis of stained slides;
providing the digital image as an input to the QA machine learning model based on the quality designation being an approved quality designation;
receiving the disease designation for the digital image as an output from the QA machine learning model;
receiving an external designation for one of the digital image or the target specimen, the external designation generated by a user and comprising a disease property selected from at least one of a cancer detection, a cancer grade, a cancer origin, a diagnosis, a presence or absence of a microorganism, a specimen type, a cancer type, a cancer status, a tumor size, a lesions risk level, or a grade;
evaluating the external designation by comparing the disease designation to the external designation;
rejecting the external designation based on the evaluation, based on the external designation deviating from the disease designation; and
outputting a comparison result based on evaluating the external designation by comparing the disease designation to the external designation.
27. Independent Claim 7 is very similar, and primarily differs by omitting the QC machine learning model features. Claim 18 is directed to a system which uses memory and a processor to implement the method of Claim 7.
The Skilled Addressee
28. The specification as a whole only briefly mentions particular classes of machine learning algorithms. It contains no detail on how to build a machine learning algorithm. Rather it discusses what data to put into the algorithms. It is immediately apparent that the skilled addressee must be someone who is extremely familiar with how to program and implement a wide range of machine learning models, and that all forms of machine learning must be part of their common general knowledge.
29. The skilled addressee also needs to be very familiar with image processing techniques and how to incorporate these within machine learning algorithms. The specification is bereft of detail on how these operate and how to incorporate them into the workflow. Again, knowledge of these must form part of the skilled addressee’s knowledge base. Experience in medical imaging would be beneficial for supplying/creating the training data but is not essential. It is noted that the specification states that images stored by a pathologist and associated with the requisite input data and outcomes may be used for the training data. Simple access to these files would likely be sufficient for the skilled addressee.
30. I therefore consider that the skilled addressee is a programmer proficient in the implementation of machine learning algorithms, preferably with experience in image analysis related to medical (pathology) or similar images.
Construction
31. The principles underpinning construction are well-established. The following summary was provided in Austal Ships:[24]
[24] Austal Ships Sales Pty Ltd v Stena Rederi Aktiebolag [2008] FCAFC 121 (“Austal Ships”) at [13]-[14]
“In Flexible Steel Lacing Company v Beltreco Ltd [2000] FCA 890; (2000) 49 IPR 331, Hely J considered at length the approach to construction of a specification and, in particular, the circumstances in which uncertainty might lead to invalidity. At [71]-[78] his Honour identified the following principles:
·The monopoly must be defined in a way that is not reasonably capable of being misunderstood.
·In determining the nature and extent of the monopoly claimed, the specification must be read as a whole, but recognizing that the parts have different functions. The claims mark out the legal limits of monopoly. What is not claimed is disclaimed. The specification describes how to carry out the process and the best method known to the patentee of doing so.
·Although the claims are construed in the context of the specification as a whole, it is not legitimate to narrow or expand the boundaries of the monopoly as fixed by a claim by adding glosses drawn from other parts of the specification. If a claim is clear, it is not to be varied, qualified or made obscure by statements found elsewhere in the document.
·It is legitimate to refer to the rest of the specification to explain the background to the claims, to ascertain the meaning of technical terms and resolve ambiguities in the construction of the claims. When the language of the claims is obscure or doubtful such doubts may be resolved by reference to the specification.
·It is not necessary that the claims be construed without reference to the body of the specification in order to see whether there is any ambiguity. The document is construed as a whole. If the specification demonstrates an intention that words used elsewhere have a particular meaning, effect should be given to such a ‘dictionary’.
At [79]-[81] his Honour then continued:…
[81] Other principles of construction which may be of assistance in the resolution of the present matter include:
· A patent specification should be given a purposive construction rather than a purely literal one…
· The hypothetical addressee of the patent specification is the non-inventive person skilled in the art before the priority date. The words used in a specification are to be given the meaning which the hypothetical addressee would attach to them, both in the light of his own general knowledge and in the light of what is disclosed in the body of the specification.
· There is a fine line between, on the one hand, reading down the words of a patent claim to reflect how a person skilled in the art would understand it in a practical and commonsense way, and, on the other hand, impermissibly limiting the clear words of a claim because a reader skilled in the art would be likely to apply those wide words only in a limited range of all the situations they describe.
· It is permissible for an invention to be described in a way which involves matters of degree. Lack of precise definition in claims is not fatal to their validity, so long as they provide a workable standard suitable to the intended use. The consideration is whether, on any reasonable view, the claim has meaning. In determining this, the expressions in question must be understood in a practical, commonsense manner. Absurd constructions should be avoided and mere technicalities should not defeat the grant of protection.
· As a general rule, the terms of a specification should be accorded their ordinary English meaning.
· Evidence can be given by experts on the meaning which those skilled in the art would give to technical or scientific terms and phrases and on unusual or special meanings given by such persons to words which might otherwise bear their ordinary meaning.
· However, the construction of the specification is for the court, not for the expert witness. In so far as a view expressed by an expert depends upon a reading of the patent, it cannot carry the day unless the court reads the patent in the same way.
· Section 116 of the 1990 Act provides that the court may, in interpreting a complete specification, refer to the specification without amendment. However, it is neither useful nor legitimate to do so where the amended specification is clear.”
32. The claims do not possess any major issues in terms of clarity. However, it is useful to highlight some of the key terms here.
“a target specimen associated with a pathology category”
33. “A pathology category” refers to a general class of work within pathology. Some example categories are given in the specification, including: histology, cytology, frozen section, and immunohistochemistry.[25] Any given subfield within pathology where specimens are analysed would be appropriate. While the “associated with” portion indicates any and every potential relationship between a specimen and the pathology category could be considered, a purposive construction limiting the specimen to one which has undergone sample preparation/analysis from said pathology category is more appropriate.
[25] Specification at [034], [056]
Approved quality designation
34. As defined by Claim 1, the quality designation is the output of the QC machine learning model. An ordinary everyday meaning would define this feature as some form of evaluation of the quality of the specimen. It could be as simple as pass/fail, contain a checklist of artifacts present/not in the specimen, or some form of ranking system. The specification[26] also indicates that an everyday meaning is appropriate for the term “quality designation”.
[26] Ibid at [033], [054], [064]-[065]
35. When preceded by the word “approved” at least two options become apparent: either approved can refer to the specimen being analysed or it can refer to the output of the machine learning model being of a type which is known to the system. I note that the specification states that “The quality designation may be an approval or a rejection and may also include a scale”.[27] In later paragraphs it is stated that the quality designation may need to meet a threshold.[28] Thus it becomes clear that an “approved quality designation” is meant to refer to the specimen meeting some form of quality threshold.
[27] Ibid at [054]
[28] Ibid at [0101]-[0102]
Disease designation and External designation
36. Claims 1, 7, and 18 define “disease designation” to be the output of the respective QA machine learning models. An ordinary meaning would indicate that the result of the analysis has identified that some form of disease is present or absent from the specimen. It could include quantified measures or qualitative ones. The specification describes it as “A disease designation may be one or more of cancer detection, cancer grade, cancer origin, diagnosis, a presence or absence of a microorganism, specimen type, cancer type, cancer status, tumor size, lesions risk level, grade, or the like.”[29] It further states that diseases other than cancer may also be diagnosed.[30] It is therefore apparent that an ordinary meaning is applicable for this term.
[29] Ibid at [079]
[30] Ibid at [040]
37. The external designation contains the same type of information as the disease designation.[31] It differs only in that the designation was not performed by the QA machine learning model. Rather, it is supplied by a pathologist or by some other user.
Manner of Manufacture
[31] Ibid at [080]
Applicable Law
38. Section 18 of the Patents Act 1990 provides that:
(1) Subject to subsection (2), an invention is a patentable invention for the purposes of a standard patent if the invention, so far as claimed in any claim:
(a) is a manner of manufacture within the meaning of section 6 of the Statute of Monopolies;
...
39. The High Court considered whether a method for eradicating weeds from crop areas possessed a manner of manufacture in NRDC[32]
“ … The right question is: ‘Is this a proper subject of letters patent according to the principles which have been developed for the application of s. 6 of the Statute of Monopolies?’ ”
[32] National Research Development Corporation v Commissioner of Patents [1959] HCA 67 (“NRDC”) [14] at 269
40. NRDC also explained that[33]
“ … The point is that a process, to fall within the limits of patentability which the context of the Statute of Monopolies has supplied, must be one that offers some advantage which is material, in the sense that the process belongs to a useful art as distinct from a fine art … - that its value to the country is in the field of economic endeavour. … ”
[33] NRDC [22] at 275, (citations removed)
41. Myriad[34] qualified that the NRDC ruling was not a strict formula to be applied to determine whether an invention was patentable, rather a case-by-case analysis was required:
“This Court in NRDC did not prescribe a well-defined pathway for the development of the concept of ‘manner of manufacture’ in its application to unimagined technologies with unimagined characteristics and implications. Rather, it authorised a case-by-case methodology. Consistently with that approach, and without resort to the ‘generally inconvenient’ proviso in s 6 of the Statute of Monopolies, there may be cases in which the court will decide that the implications of patentability of a new class of invention are such that the invention as claimed should not be treated as patentable by judicial decision.”
[34] D’Arcy v Myriad Genetics Inc. [2015] HCA 35 (“Myriad”) at [23]
42. In recent times a number of decisions have issued which are related to whether various computer-related inventions satisfy the requirements of s 18(1)(a), including Research Affiliates[35] and RPL[36]. The claims directly reference “A computer-implemented method”, as well as numerous software-related functions. These decisions therefore are highly relevant to the present case.
[35] Research Affiliates LLC v Commissioner of Patents [2014] FCAFC 150 (“Research Affiliates”)
[36] Commissioner of Patents v RPL Central Pty. Ltd. [2015] FCAFC 177 (“RPL”)
43. RPL identifies that:[37]
“A claimed invention must be examined to ascertain whether it is in substance a scheme or plan or whether it can broadly be described as an improvement in computer technology. The basis for the analysis starts with the fact that a business method, or mere scheme, is not, per se, patentable. The fact that it is a scheme or business method does not exclude it from properly being the subject of letters patent, but it must be more than that. There must be more than an abstract idea; it must involve the creation of an artificial state of affairs where the computer is integral to the invention, rather than a mere tool in which the invention is performed. Where the claimed invention is to a computerised business method, the invention must lie in that computerisation. It is not a patentable invention simply to 'put' a business method 'into' a computer to implement the business method using the computer for its well- known and understood functions.
Is the mere implementation of an abstract idea in a well-known machine sufficient to render patentable subject matter? Is the artificial effect that arises, because information is stored in RAM and there is communication over the Internet or wifi, sufficient? Does any physical effect give rise to a manner of manufacture? Are the mere presence of an artificial effect and economic utility, without more, sufficient to determine manner of manufacture?
It is not a question of stating precise guidelines but of deciding, in each case, whether the claimed invention, as a matter of substance not form, is properly the subject of a patent.”
[37] RPL at [96]-[98], (emphasis added)
44. RPL further includes:[38]
“Recognising that the claims are to a method and system comprising a combination of integers, it is necessary to understand where the inventiveness or ingenuity is said to lie. Turning to the integers of the invention … it is apparent that, other than the integers providing that the computer processes the criteria to generate corresponding questions and presents those questions to the user, the method does not include any steps that are outside the normal use of a computer.”
[38] RPL at [112], (emphasis added)
45. Research Affiliates[39] also identifies, for the purposes of assessing manner of manufacture, that:
“… there is a distinction, between mere implementation of an abstract idea in a computer and implementation of an abstract idea in a computer that creates an improvement in the computer”.
[39] Research Affiliates at [103]
46. The considerations used to assess the substance of the invention from RPL, Research Affiliates, and others, were summarised in Aristocrat ’16 by the delegate:[40]
[40] Aristocrat Technologies Australia Pty Limited [2016] APO 49 (“Aristocrat ‘16”) at 35
“I conclude that it is relevant to consider a range of matters. Without seeking to be exhaustive, these include:
·there must be more than an abstract idea, mere scheme or mere intellectual information;
·is the contribution of the claimed invention technical in nature;
·does the invention solve a technical problem within the computer or outside the computer;
·does the invention result in improvement in the functioning of the computer, irrespective of the data being processed;
·does the application of the method produce a practical and useful result;
·can it be broadly described as an improvement in computer technology;
·does the method merely require generic computer implementation;
·is the computer merely an intermediary or tool for performing the method while adding nothing of substance to the idea;
·is there ingenuity in the way in which the computer is utilised;
·does the invention involve steps that are foreign to the normal use of computers; and
·does the invention lie in the generation, presentation or arrangement of intellectual information.”
47. Many of the above points have been used in several examination reports, responses, and in the Applicant’s submissions for this decision. Not all of the points are relevant to this case, and the list is indicative only of factors which may need to be considered. They are not meant to be used as a box-checking exercise – as per Myriad, a case-by-case approach is required.
The Examiner’s Reports
48. The Examiner initially characterised the problem and solution in the following manner.[41]
[41] Examination Report No 1
“The problem addressed by the present invention is understood to improve existing systematic quality assurance methods, which currently are impractical and inefficient due to the inclusion of multiple pathologists for review, which is resource and time intensive.
The invention addresses this problem through inputting digital pathology images into machine learning algorithms and compare the quality of the output to that of a professional diagnosis.
However, it is clear from reading the specification as a whole that the claimed invention merely makes use of known technology, such as use of machine learning models for processing images and generating a quality parameterised output. These technologies are utilised in the conventional way, in order to reduce labour cost of trained health professionals.
Given the above, I consider that the substance of the invention is the machine learning algorithms defined at the level of detail defined in the claims, rather than any specific aspect of how these algorithmic operations are implemented by a computer. It is clear that there has not been a technical improvement to the operation of these elements, rather it is apparent that each integer of the claimed invention operates in its normal and expected way and has merely been collocated in the present invention. This arrangement does not result in a technical improvement to any of the constituent elements, nor does a technical improvement result from the collocation as a whole. Therefore, it cannot be said that the invention provides a technical solution to the problem at hand.
The invention is therefore in substance, merely a digital image processing scheme to provide quality or disease parameters of pathology images.
…
The claims define the use of a computer to receive a digital image of the specimen and use a machine learning model for quality control/assurance in determining an indication of disease present within the sample or the image’s quality itself. The technical aspects that are being defined by the claims are performed using only routine operations of a computer and therefore provide no technical advantage over what are already well-known functions of a computer. Hence, the claims define an image processing scheme merely implemented by a conventional computer.”
49. Further examination reports issued. The main source of contention was whether the invention was technical in nature or merely a scheme. Several of these further reports used headings which correspond to those outlined in Aristocrat ’16. Helpfully, the Applicant has also used the same headings in their submissions for this decision. I likewise use those same headings.
Consideration
Does the invention solve a technical problem?
50. The Examiner considered that the problem addressed by the current application could be identified from paragraphs [002]-[003] of the specification. These paragraphs outline that the quality control and quality assurance processes within a pathology laboratory are impractical and inefficient as they require the services of two or more pathologists. The Examiner stated that:[42]
“Addressing the conventional resource and time-intensive methods conducted by pathologists, is an administrative problem, rather than a problem which is technical in nature. From this passage, it is clear that the disclosed problem does not relate to a technical deficiency, but rather inefficiencies in current quality control and quality assurance procedures. Such a problem cannot be considered technical in nature.”
[42] Examination Report No 3, p4
51. Following the proposed amendments the Applicant submits that: [43]
“Yes, the invention addresses the technical problem of identifying biomarkers using stained slides. The invention is capable of determining biomarkers that would not usually be readily identifiable using stained slides and would instead require additional testing, as discussed above[44]. In this way, the invention solves a technical problem.”
[43] AWS at [059]
[44] AWS at [054], citing specification at [029]
52. Without an explicit indication of which paragraphs the Applicant means by “as discussed above” I will try to infer their intent. AWS paragraphs [051]-[057] briefly discuss the biomarkers and cite excerpts from the specification at [028] and [029]. Apart from these paragraphs there is nothing else to support the Applicant’s assertion in the AWS or in previous responses to the various Examination Reports.
53. It is useful to refer to the specification indicated by the Applicant. Paragraph [028] briefly describes that H&E-stained histologic preparations do not provide sufficient information for a pathologist to visually identify biomarkers that can aid diagnosis or guide treatment in every case. [028] also states that if H&E staining does not identify a potential diagnosis a pathologist would then use additional techniques, such as immunohistochemistry, immunofluorescence, in situ hybridization, fluorescence in situ hybridization, or genetic testing.
54. This is followed by [029] which indicates that a pathologist will obtain a digitised image of a stained microscope slide (containing a specimen) and estimate the number of abnormal stained cells. It is stated that this is time consuming and may lead to errors. While [029] does briefly mention an example where AI, using a H&E-stained image, can be used to infer the presence of biomarkers which would ordinarily require at pathologist to perform additional tests to verify, the specification does not disclose any further information than this.
55. I further note that paragraph [026] includes:
“Thus, the process of obtaining stained slides and tests may be done automatically before being reviewed by a pathologist. When paired with automatic slide quality review and result determination, this may provide a fully automated slide preparation and evaluation pipeline in parallel with a pathologist review.”
56. As far as identifying the problem to be solved, I am not persuaded by the Applicant’s submissions as they primarily refer to a potential outcome. From my reading of the specification, as outlined above at [7]-[25], it is abundantly clear that the majority of the specification is directed towards the use of machine learning models to perform QC and QA tasks instead of having additional pathologists or technicians perform them. After considering the entire specification it is my view that the problem the specification seeks to address is how can the number of pathologists required to provide quality control and quality assurance of pathology specimens be reduced. The problem of choosing not to hire additional pathologists to perform QC or QA is a business or administrative decision and I do not consider that the problem itself is technical in nature. This does not automatically preclude an invention from being patentable – rather, to be patentable, some portion of the solution needs to utilise a technical modification or a technical adaption to some form of technology.
Does the invention result in an improvement in the functioning of the computer, irrespective of the data being processed? Does the method merely require generic computer implementation? Can it be broadly described as an improvement in computer technology?
57. On these points the Applicant only submitted:[45]
“Yes, the invention provides an improvement to pathology computing devices by enabling the determination of a disease designation using biomarkers that would not have previously been readily identifiable in stained thin section. The invention also does not rely on generic computer implementation, as the machine learning models are specifically adapted to perform the invention.”
[45] AWS at [062], emphasis in original
58. Addressing the Applicant’s latter assertion initially, the specification[46] discloses that the invention can be performed “through one or more computers, servers, and/or handheld mobile devices”. The specification does not describe any hardware improvements, modifications, or adaptations which are required to implement the QC or QA machine learning models. As described in the specification, large families of machine learning models are contemplated.[47] There is no improvement to any specific machine learning model, rather an existing model is trained using various data inputs. The specification also states that the operating systems and hardware are conventional in nature.[48] From these factors I can only conclude that the software and hardware simply require standard computational resources. There is nothing beyond the routine being done to either the software or hardware, alone or in combination, and neither appear to be modified or adapted in a technical sense in order to utilise the desired training data.
[46] Specification at [042], [054]
[47] Specification at [061]
[48] Specification at [0122]
59. It is also apparent that the improvement the Applicant identifies is in relation to the desired result of the data processed by the computer. The computer does not operate faster, consume less electricity, or perform any other function which it could not already do. Instead, it processes the data using known machine learning techniques to obtain a result, that result being a quality designation or a disease designation. This is not an improvement in the computer’s hardware or software.
60. The Applicant asserts that the machine learning models are specifically adapted to perform the invention. However, it is apparent that the training, and the subsequent result of the invention, is entirely dependent on the data which is input into the machine learning models to train them. This is the only aspect of the proposed invention which distinguishes the QC and QA machine learning models from any of the generic (untrained) machine learning models mentioned. The invention cannot be said to be “irrespective of the data being processed”.
61. Thus, I consider that the invention uses generic computing hardware to train and subsequently implement a non-specified type of machine learning model. The machine learning models themselves, prior to training, are certainly generic. The invention appears to reside in the data which is used to train the machine learning models, and the subsequent use of the trained machine learning model to evaluate images of various specimens.
Is the contribution of the claimed invention technical in nature?
62. As noted in paragraph 4, the Applicant was invited to supply additional information on this item. They submitted:[49]
“Yes, by training the machine learning model using biomarkers that are indicative of one or more of an over-expression of a protein, a gene product, an amplification of a specific gene, or a mutation of a specific gene, the invention is leveraging the training data to identify or otherwise determine biomarkers of the digital images that would have otherwise required the performance, including time intensive slide preparation, of additional tests. In this way, the invention provides for a technical solution, by way of the specifically trained machine learning model, to the technical problem of how to identify biomarkers without having to perform additional testing.”
[49] AWS at [061] following correction of 15 May 2024
63. The Applicant’s position is clear. They submit that the solution which has been adopted is technical in nature, and they also assert that the problem addressed is technical. As described above, I do not agree that the problem is technical in nature. But it is still possible that a technical solution has been implemented to solve the problem at hand. It is therefore logical to evaluate whether this submission from the Applicant bears up to scrutiny or whether it is merely a phantasm which has been proffered as being real.
64. From my reading of the entire specification, the only information regarding the Applicant’s position appears in paragraphs [028]-[029]. Paragraph [028] discusses that H&E testing is often used by a pathologist, but that further types of stains or techniques can be employed if the results do “not provide sufficient information for a pathologist to visually identify biomarkers that can aid in diagnosis or treatment.” Paragraph [029] explicitly states (emphasis added):
“A digitized image may be prepared to show a stained microscope slide, which may allow a pathologist to manually view the image on a slide and estimate a number of stained abnormal cells in the image. However, this process may be time consuming and may lead to errors in identifying abnormalities because some abnormalities are difficult to detect. Computational processes using machine learning models and devices may be used to assist pathologists in detecting abnormalities that may otherwise be difficult to detect. For example, AI may be used to predict biomarkers (such as the over-expression of a protein and/or gene product, amplification, or mutations of specific genes) from salient regions within digital images of tissues stained using H&E and other dye-based methods. The images of the tissues could be whole slide images (WSI), images of tissue cores within microarrays or selected areas of interest within a tissue section. Using staining methods like H&E, these biomarkers may be difficult for humans to visually detect or quantify without the aid of additional testing. Using AI to infer these biomarkers from digital images of tissues has the potential to improve patient care, while also being faster and less expensive.”
65. The wording of the emphasised portions utilises the terms “may” and “has the potential”. Nothing definite is described. The text does not indicate that the proposed method will result in a machine learning model being able to detect biomarkers whereby a pathologist would ordinarily be required to perform additional testing to generate a diagnosis. For example, it does not state that by using the machine learning model to analyse H&E stained samples that tests which would ordinarily require (F)ISH, genetic testing, etc do not need to be performed following the prediction. Nor does the paragraph provide any detail on how the machine learning model would be trained such that it could identify abnormalities which would challenge a pathologist – noting that the training data for the machine learning models must have been determined by a pathologist at some stage. Rather, paragraph [029] merely states that biomarkers may be predicted in images which contain H&E and other dye-based stains.
66. Further, and as per the quoted portion in paragraph 55 above, the machine learning models are designed to operate before, after or in parallel with a pathologist review – or in the language of the claims, an external designation needs to be generated by a user. The consequence is that if the pathologist cannot visually identify the disease state using H&E testing then they would necessarily need to perform additional tests to obtain the external designation. In such a case there is no difference in the number of tests performed. It is only a matter of whether the data analysed by the machine learning models is limited to particular types of image data, such as H&E image data.
67. It follows that I do not agree with the Applicant’s argument. The contribution does not lie in a trained machine learning model being able to identify biomarkers which it could not already do.
68. In contrast, the Examiner does not consider the contribution to be technical in nature. They characterised the contribution as follows:
“… the method employs conventional computer capabilities to execute a standard process using established techniques in image analysis, machine learning, and medical diagnostics. The computer's role is pivotal for implementing operations like applying machine learning models and processing images, yet it doesn't introduce significant advancements beyond established knowledge. The method's value lies in systematically applying existing techniques to address specific challenges, making the computer an intermediary tool. The process of submitting samples for machine learning model approval doesn't notably enhance the computer's capabilities, rather it utilises standard computational tools to analyse and convey image information, without fundamentally improving the computer's core functions.”
and
“… The invention lies in the two different non-specific machine learning models to perform different tasks. These machine learning models are combined merely in the fact that the output of one model is provided as input of the secondary model. These models are used to address administrative inefficiencies within pathological image analysis and leverage standard techniques and technologies to solve the above problem.”
69. Based on the level of detail disclosed by the specification for the image analysis and machine learning models, these must be established techniques. The specification does not disclose any information regarding any processes a pathologist would use to review pathology images, merely that they routinely do so. For example, it is unknown whether the pathologist would ordinarily use machine vision techniques to alter contrast, colour balance, perform counting of cells, etc. In the absence of any information it may only be assumed that the processes are entirely standard and any contribution lies elsewhere.
70. In relation to the training of the different machine learning models, it is apparent that the training of the machine learning models must also be well known in the art. The specification provides no information beyond stating that QC or QA training images are used as input, along with various QC or QA outputs.[50] While the amended claims and above paragraphs contain some brief information regarding the what the training images contain (namely that some samples are normal, contain missing tissue, bubbles, over-expression of a gene, etc), there is no information beyond the machine learning models being supplied with the types of errors or diseases which they are designed to detect. Whatever steps are necessary to implement the machine learning models themselves are left entirely to the skilled addressee. There is no contribution to how the machine learning model operates. Any contribution in relation to the machine learning models is limited to the types of data processed by the machine learning models.
[50] Description at [054], [059]-[061] and [071], [074]-[077]
71. The Examiner stated that the output of the one model (the QC machine learning model) is merely provided as input to the second model (the QA machine learning model). I agree that Claim 1’s feature of “providing the digital image as an input to the QA machine learning model based on the quality designation being an approved quality designation” simply requires that the output of the QC machine learning model is used as a filter, weighting function, or similar based on the specimen’s results after passing through the QC machine learning model. The claim does not address what happens for specimens that do not meet the approved quality designation.
72. It is apparent that the claims do not define any technical contribution in how the models are combined beyond a trivial, logical sequencing. The specification likewise provides no further information which would indicate some form of technical modification occurs as a result of performing the computational analysis in this particular order.
73. The final portion of the claimed invention requires an external designation be provided, where the external designation is obtained via a user and any method, excluding the utilisation of the QA machine learning model, may be implemented. There is a dearth of information on how the external designation is obtained, merely that it is. It cannot be considered a technical modification or adaptation. There is also nothing technical in simply comparing the information supplied from the QA machine learning model with that supplied from the user. The mere comparison of similar types of data is ubiquitous in the computational and numerous other arts. Noting that functional claiming has been employed, every possible method to compare the two sets of data must be known by the skilled addressee. It follows that this step cannot be considered a technical modification or technical adaptation.
74. When the features related to obtaining the external designation are combined with the features related to obtaining a QA designation it becomes clear that the Applicant’s argument that additional testing can be avoided is not persuasive. In the event where the QA designation provides additional or different information in advance of obtaining the external designation, it might lead the external designator to take further care to analyse the images or might cause them to run additional tests to verify whether the QA machine learning model’s prediction is correct or not. This does not obviate or reduce the need for further testing. At best it might reduce the time required to obtain a diagnosis. Where the QA is run in parallel or subsequent to analysis by the external designator there is no time benefit.
75. From all of the above it becomes apparent that any contribution lies in the data used to train the QC and QA machine learning models, the sequencing of the machine learning models, and the use of the machine learning models to process future pathology samples. The mere processing of data to obtain a desired result, where the implementation is effectively at the whims of the skilled addressee, is simply using well known and well understood computer technology for its known purpose. It is not technical in nature. Similarly, the sequence of providing of one model ahead of another does not involve any technical aspects and cannot be said to be technical in nature. There are no technical contributions from any of the computing aspects of the invention.
76. The Applicant has also made submissions stating that the present invention is an advance to the general field of pathology.[51] However, as above, the output of the QA machine learning model needs to be compared against the work of a pathologist. While two pathologists may ultimately be reduced to one, this is an administrative choice. The invention does not necessarily improve the quality of the diagnosis and does not provide for results which could not be obtained before. I do not see how this can be an improvement to the field as a whole.
[51] AWS at [010]
77. The Applicant has further submitted that the present invention recites a reduction to practice, and that the substance of the invention has been stripped of all technical character.[52] Several paragraphs are then devoted to arguing that the training of the machine learning models must use data which contain the various results. For example, that the QC machine learning model must be trained with sample images which contain unwanted artefacts such as a missing tissue, a glass crack, a bubble, a blur amount, a missing tissue, a folded tissue, a line, a scratch, dust, pen, or a stain amount. Similarly, the QA machine learning model must be trained using data where biomarkers indicate particular disease states.[53]
[52] AWS at [037]-[039]
[53] AWS at [043]-[057]
78. The challenge I see is that the specification only describes the desired input training data and desired results. It cannot be a reduction to practice if every possible way to implement the result is left to the skilled addressee. I am reminded by Allsop J et al. in Encompass Corporation Pty Ltd v InfoTrack Pty Ltd [2019] FCAFC 161 at [100]-[101] where a similar submission was determined to merely be the idea of implementing the desired software, noting that the specification in that case was devoid of any particular software or algorithms. The same applies here. All that is described is the desired result, that being a trained machine learning model which can determine whether output the types of data used in the training of the machine are present in future samples.
79. I can only conclude that the contribution of the proposed invention is not technical in nature.
Does the application of the method produce a practical and useful result?
80. The Applicant has argued that the end result is both practical and useful in that (i) the amount of data reviewed by a pathologist which does not conform to the quality metric is minimised, and that (ii) biomarkers which are not readily identifiable by a pathologist may now be detected.[54] The Applicant has cited portions of Bio-Rad Laboratories, Inc. [2018] APO 24 (Bio-Rad) to support their contention that the claimed invention has utility outside of the computer and “it is unnecessary to separately consider any technical improvement to the computer technology itself.”[55]
[54] AWS at [065]-[067]
[55] AWS at [066]-[067], which cites Bio-Rad at [57]
81. Bio-Rad uses the same principles derived from the case law which I have used – it does not set any form of binding precedent or represent anything more than a decision by the delegate on the merits of the case before them. I agree with the Applicant’s submission that when the method of Bio-Rad is implemented a quality control strategy is selected which minimises the total number of samples (destructively) tested to ensure their medical apparatus operated within the necessary calibration limits. However, this does not alter the facts of the present application.
82. The present invention does not state anywhere that the QC machine learning model prevents the pathologist from reviewing a sample which has failed to meet the desired QC metric. The specification states “In use, the disclosed systems may perform QA/QC analysis and delivery before or after a received diagnosis (e.g. by a pathologist)”.[56] Thus, it would be mere happenstance if the system were programmed to prevent the pathologist analysing a sample which the QC machine learning model had flagged as not-approved. Further, this depends entirely on when the pathologist performs their checks compared to when the QC and QA machine learning models process the specimen data.
[56] Specification at [039]
83. While the specification also states that a notification may be generated and sent to an “appropriate health care professional, such as a histology technician” to prepare another specimen, or to the scanner operator to rescan the specimen,[57] again there is no mention in the specification as filed that the pathologist does not view specimens which fail the QC metric. I fail to see how a technical result, such as the minimisation of destructive testing, necessarily occurs when the present application is put into practice.
[57] AWS at [0102]
84. The output of the QA machine learning model is only a prediction or an inference that biomarkers are present within a specimen. It does not definitely identify the biomarkers. A pathologist is still required to independently perform their analysis to verify whether the prediction is correct or not. Thus, while the output data is useful in the sense that it provides further information to a pathologist, and the data is practical in that it can aid in providing a diagnosis, it does not necessarily give rise to a technical effect. That is, the information provided to the pathologist is merely there to aid them in making a decision on what diagnosis to give. The making of a decision is an abstract concept.
85. Regarding the Applicant’s second contention, for the reasons given in the section above, I am not convinced biomarkers which were not readily identifiable are now able to be detected.
Is the computer merely an intermediary or tool for performing the method while adding nothing of substance to the idea? Is there ingenuity in the way in which the computer is utilised?
86. There is no question that the computer is essential to perform the invention – a machine learning model cannot operate without a computer. However, the computer itself is of no significance other than it being an object on which the necessary processing is performed. The computer is only an intermediary.
87. As described above, a machine learning model is supplied with QC or QA input training data and QC or QA output training data in order to train it to become a QC or QA machine learning model. It is only the data used to train the models which differs from what is known in the art, and then the subsequent use of the trained machine learning models to replicate the disease identification work of a pathologist. Nothing has been described which indicates there is ingenuity in the way in which the computer is used. The implementation, or the “how”, is left entirely to the skilled addressee.
88. The result is that any ingenuity lies in the idea to use pathology samples to train the QC or QA machine learning models, and then use these trained machine learning models to replicate the work of a pathologist. While the submissions frequently referred to the machine learning models being able to detect biomarkers which a pathologist could not, there is no detail in the specification regarding how this occurs. Any ingenuity lies in the idea of performing these actions instead of how it is implemented. Should there be any ingenuity in how the machine learning models detect biomarkers which a pathologist cannot then, as the specification is silent on these factors, further consideration of this would need to occur on s40 grounds.
Does the invention lie in the generation, presentation or arrangement of intellectual information?
89. While the Applicant has not addressed this heading directly, the submissions which are directed to whether the invention is practical and useful also contain submissions which indicate that the Applicant considers the invention is more than merely the processing of information.
90. In the present invention it is clear that training data is input into a machine learning model to train it. The trained model is then used with subsequent data to identify different quality states or diseases. It is plain that the invention lies in the arrangement and generation of information. The information obtained is not utilised in a technical manner after it has been determined. For example, it is not used as an input to trigger the rescanning of a sample via an automatic slide scanner. Rather, the information obtained is merely compared against what the pathologist has done as part of the external designation. In this respect I cannot see how the information can be anything other than intellectual in nature.
Summary on Manner of Manufacture
91. As I identified above the contribution lies in the data used to train the QC and QA machine learning models, the sequencing of the machine learning models, and the use of the machine learning models to process future pathology samples. The substance lies in the administrative scheme to automate the role of a pathologist using machine learning models which have been trained using particular types of pathology image data. The end result is that only one pathologist is needed to process samples instead of two or more pathologists.
92. There is nothing technical in the implementation of the above administrative choice. While the results of the processed data is useful, this is only as a result of the administrative scheme. It follows that Claim 1 does not satisfy the criteria of s18(1)(a). Claims 7 and 18 are broader than Claim 1 and omit the QC elements. They fail to satisfy s18(1)(a) for the same reasons as Claim 1.
93. I have also considered each of the dependent claims. There are no details present which would indicate a technical advance has occurred in any form of technology. They merely describe further desired results of the present invention. It follows that none of the claims satisfy s18(1)(a).
Other matters - Section 40
Legal Principles
94. Section 40(2)(a) of the Act requires that the complete specification must disclose the invention in a manner which is clear enough and complete enough for the invention to be performed by a person skilled in the relevant art, and s40(3) requires that the claims be clear and succinct and supported by matter disclosed in the specification.
95. In CSR Building Products Limited v United States Gypsum Company[58], a Deputy Commissioner determined that the steps involved in assessing whether the disclosure requirement is satisfied are:
(i)Construe the claims to determine the scope of the invention as claimed;
(ii)Construe the description to determine what it discloses to the person skilled in the art; and
(iii)Decide whether the specification provides an enabling disclosure of all the things that fall within the scope of the claims.”[59]
[58] CSR Building Products Limited v United States Gypsum Company [2015] APO 72 (CSR)
[59] CSR at [95]
96. In Evolva SA[60], a Deputy Commissioner reformulated the third question stated in CSR as a two-step consideration:
(a)Is it plausible that the invention can be worked across the full scope of the claim?
(b)Can the invention be performed across the full scope of the claim without undue burden?[61]
[60] Evolva SA [2017] APO 57 (Evolva)
[61] Evolva at [45].
97. The approaches taken in CSR and Evolva found approval with the Federal Court in Cytec Industries Inc. v Nalco Company[62].
[62] Cytec Industries Inc. v Nalco Company [2021] FCA 970 (Cytec) at [143]-[149]; 162 IPR 202.
Brief considerations – undue burden
98. While not raised by the Examiner, in writing the above decision it is apparent that the broad functional claiming and limited description of anything other than the desired result has left the bulk of the invention in the hands of the skilled addressee. For example, the skilled addressee needs to identify the necessary training data, identify the type of machine learning model, program it in its entirety, identify what types of further image processing might be required, perform any necessary model supervision, and there are likely to be many other significant factors which have been left entirely to the skilled addressee.
99. Apart from the general idea of the invention, the skilled addressee is effectively left to their own devices. This could indicate the presence of an undue burden. However, noting that the Applicant has not had the opportunity to comment on this aspect, and that providing such an opportunity would not serve any benefit to the overall outcome of this decision, I elect to not delve any further or make any particular finding on this issue.
Conclusions
As identified above the claimed invention does not satisfy the criteria to be a manner of manufacture. I have also read and analysed the entire specification. I do not see any material which could be added which would overcome this finding. It is therefore appropriate that I refuse the application.
Dr David Carberry
Delegate of the Commissioner of Patents
0
12
0