Publication: Artificial Intelligence (AI) in Pathology – A Summary and Challenges Part 4

6. AI – issues to be resolved

Over the last 100 years both The Covid19 and The Spanish Flu pandemics have shown their disproportionate impacts on patients of low income and racial minorities. A combination of diagnostics bias and sample bias have been the culprit for the global healthcare disparities. Evans argues that present diagnostic tools often fail patients who do not fit the prospects of the majority.⁵⁸ Even though there is an active effort to involve females in clinical study samples there are many treatment and drug advices that are founded on findings taken from the samples of Caucasian males. The author proposes, going forward, to decode the present and reshape existing practices before implementing AI to avoid existing biases and further increasing health disparities.⁵⁹

Colling et al. propose a UK‐wide strategy for AI and DP. If the requirements of proper slide image management software, integrated reporting systems, improved scanning speeds, and high-quality images for DP systems are achieved then it will provide time and cost saving benefits over the traditional microscope based pathology approach and reduce problem of inter‐observer variation. The successful introduction of AI and DP tools to the healthcare system will need proper regulatory approved and evidence based validation, and a lowering of the resistance to collaborate between academic and industry developers.⁶⁰ Robertson et al.’s work discusses the limitations of deep learning as it works well in supervised learning but not for unsupervised learning. The deep learning approach is not suitable for the discovery of novel biomarkers, as it being an unsupervised learning problem. If the model is educated only by means of images attained from imaging equipment from a single merchant then it may fail to react acceptably to images acquired from the equipment of another merchant. They observe the challenges to having a full digital workflow, a must for deep learning, due to the high costs and the dependence on solid IT support systems.⁶¹ Typically, training of DL models requires many of annotated samples that belong to dissimilar categories. However, in reality it can be hard to collect a balanced dataset for training because of the fact that certain ailments have a low prevalence causing problem of data. Studies have shown that many models that perform well on balanced datasets do not when it comes to their imbalanced counterparts.⁶² Most of medical image datasets possess this imbalance problem. One-class classification, which emphases on learning a model using examples from only a single given class, is used as an approach to overcome the problem of imbalance. Gao et al. proposed a novel method which allows DL models to leverage the concept of imaging complexity to optimally learn single-class-relevant inherent imaging features. They then compared the effects of perturbing operations used on images to realize imaging intricacy to boost feature learning, and allowing their method outperforming four advanced methods.⁶³

Tizhoosh et al. explore problems that must be solved in order to exploit opportunities for the AI promises in computational pathology. The challenges discussed include: i) Lack of labeled or annotated data can be overcome by using active learning applied to labeling with public datasets, ii) Pervasive variability: infinite number of patterns due to presence of several tissue types (connective tissue, nervous tissue, epithelium, and muscle) required by AI algorithms to be learned, iii) Non-Boolean nature of diagnostic tasks as binary language of ‘yes’ or ‘no’  can be possible in only easy pathological cases but is rarity in the clinical practice, iv) Dimensionality obstacle:  Use of “Patching” (divide an image into small tiles) as WSI sizes typically are larger than 50K x 50K pixels, v) Turing test dilemma: A machine can be as intelligent as a human and Turing test for DP is explicitly not known, vi) Uni-task orientation of weak artificial intelligence as Deep ANNs are designed to perform only one task requiring independently training multiple AIs for tasks of classification, segmentation, and search, vii) Affordability of required computational expenses for adoption of DP is a challenge due to high costs of acquisition and storage of gigapixel histopathological scans, viii) Adversarial attacks (Targeted manipulation of a very small number of pixels inside an image can mislead the network) as negligible presence of artifacts produce misdiagnosis, ix) Lack of transparency and interpretability which is not acceptable to the physicians as there is a lack of explanation on why AI made a specific decision in reference to histopathology scans, and x) Realism of AI as the pathology community has yet to buy in fully due to its issues related to ease of use, financial return, and trust. The authors describe multiple opportunities of: a) Deep features – Pretraining is better using Transfer learning instead of training a new network from scratch, b) Handcrafted features (such as gland shape and nuclear size) – Do not forget computer vision as it can be applied in DP to attain high identification accuracies, c) Generative frameworks: Learning to see and not judge as Generative models, focus on acquiring to reproduce data instead of making any decision such as pulmonary disease categorization and for functional MRI analysis, d) Unsupervised learning: When we do not need annotations in self-organizing plots and hierarchical clustering, and effectively combine them in the workflow of usual practice of pathology as annotating images is not portion of the everyday work of pathology experts, e) Virtual peer review – Placing the pathologist in the central to both algorithm development and execution: Algorithms extract reliable information from proven archived diagnosed cases similar to the relevant features of the patient, that are diagnosed and treated by other physicians;  Comparing for example diagnosis of patient’s cervix biopsy to a prior Pap test assessment for real-time cytologic-histopathologic correlation, f) Automation with AI can assist with case triage by performing laborious tasks for example of screening for easily identifiable cancer types or counting mitoses, and with simplification of complex tasks (e.g., triaging biopsies which require immediate action and ordering suitable stains upfront when specified); AI algorithms have attained sensitivity above 92% for breast cancer recognition, g) Re-birth of the hematoxylin and eosin image: combination of computational pathology and emerging technologies of multiplexing and three-dimensional imaging allows analysis of individual pixels of pathological images to understand diagnostic, and theoretically available prognostic information, h) Making data science accessible to pathologists will enhance their accuracy with the use of AI tools to generate/analyze big image data.⁶⁴

To integrate AI based algorithms into the workflow of pathologists, Jiang et al. outlined and discussed various challenges facing their implementation in pathology. The challenges include: i) Validation: AI models are typically established on small‐scale data and images from single‐center and therefore they need to be sufficiently validated using multi‐institutional data before clinical adoption, ii) Interpretability: DL-based AI methods are rightfully perceived as ‘black‐box’ methods due to their lack of interpretability which is an obstacle towards the clinical adoption by doctors, iii) Computing system: Histopathological photo file dimensions are typically  1,000x of an X‐ray and 100x of a CT image files requiring powerful computing and storage, and bandwidth to transmit gigapixel‐sized images, iv) Attitude of pathologists: Due to the lack of AI based model’s interpretability, pathologists are afraid of the change in workflow and worry about how to describe the evidence from AI in the diagnosis report, v) Attitude of clinicians and patients: In order to have both clinicians and patients have trust, AI based diagnostic and prognostic/predictive assays ought to have a high accuracy, and vi) Regulators: The clinical adoption of AI digital pathology needs approval by regulatory agencies and the lack of interpretability limits the approval.⁶⁵ Samek et al. present two methods that describe predictions of deep learning models to overcome DL’s black box approach. The first method which computes the sensitivity of the prediction with respect to changes in the input and the second approach meaningfully decomposes the decision in terms of the input variables.⁶⁶  Some of problems that need to be overcome to achieve the progress of DP and ML in their daily usage in pathology practice are: a) Make interfaces user friendly which currently are not, b) Require a single image format instead of current existence of several proprietary image formats, c) Overcome issue of the large image file sizes using technological advances in storing, and d)  Enhance interactions between AI experts and pathologists.⁶⁷

AI machine learning model development, a multi-step process, includes important technical, regulatory, and clinical barriers. The model should overcome these barriers, which collectively define a “translation gap,” in order to being accepted in a real world. The translation gap in digital pathology includes a variability caused by the manual nature of the tissue acquisition process and histopathology slide preparation,  differences introduced during tissue sampling, tissue fixation, sectioning, and staining. During model development and validation these variations must be accounted for in order to achieve its widespread adoption. Also, since DP is relatively immature, at present only two manufacturers have received FDA approval to market digital pathology systems for primary diagnosis.^{68 69} Similarly Steiner et al. discuss how the low penetration of digital pathology has negatively affected integration of AI into pathologist’s diagnostic workflow and validation of algorithms in live clinical settings.⁷⁰

7. Conclusion

Artificial Intelligence (AI) has come a long way over the last 65 years. Over the last two decades research in AI has gained traction in healthcare and it is now being applied across many medical subspecialties of dermatology, radiology, and pathology. A nationwide or global  strategy for AI and Digital Pathology (DP) will be necessary in order to be used for automated diagnosis, triaging cases for improved workflow, or deriving novel insights for pathologists. If DP system’s requirements of proper slide image management software, integrated reporting systems, improved scanning speeds, and high‐quality images, are achieved then it will provide time and cost saving benefits over the traditional microscope based pathology approach, offer a second opinion, and in addition it will reduce the problem of inter‐observer variation. However, AI approaches including deep learning do face rightful criticism, as their internals to make decisions by design are not known and hence will require legal and regulatory issues worked out to reap the possible benefits. The successful introduction of AI and DP tools to the healthcare system will need proper regulatory approved evidence based validation, and lowering of the resistance to collaboration between  academic and industry developers.

References

58. Evans M. “Covid’s Color Line — Infectious Disease, Inequity, and Racial Justice” N Engl J Med, 383 (July (5)) (2020), pp. 408-410, 10.1056/NEJMp2019445.
59. Straw I., “The automation of bias in medical Artificial Intelligence (AI): Decoding the past to create a better future.” Artificial Intelligence in Medicine, Volume 110, 2020, 101965. ISSN 0933-3657. https://doi.org/10.1016/j.artmed.2020.101965.
60. Colling R., Pitman H., Oien K., et al. “Artificial intelligence in digital pathology: a roadmap to routine use in clinical practice.” J Pathol. 2019 Oct;249(2):143-150. doi: 10.1002/path.5310. Epub 2019 Jul 18. PMID: 31144302.
61. Robertson S., Azizpour H., Smith K., et al. “Digital image analysis in breast pathology-from image processing techniques to artificial intelligence.” Transl Res. 2018 Apr;194:19-35. doi: 10.1016/j.trsl.2017.10.010. Epub 2017 Nov 7. PMID: 29175265.
62. Leevy J., Khoshgoftaar T., Bauder R., et al. “A survey on addressing high-class imbalance in big data.” J Big Data, 5 (2018), p. 42. https://link.springer.com/article/10.1186/s40537-018-0151-6.
63. Gao L., Zhang L., Liu C., et al. “Handling imbalanced medical image data: A deep-learning-based one-class classification approach.” Artificial Intelligence in Medicine, Volume 108, 2020, 101935. ISSN 0933-3657. https://doi.org/10.1016/j.artmed.2020.101935.
64. Tizhoosh H., Pantanowitz L. “Artificial Intelligence and Digital Pathology: Challenges and Opportunities.” J Pathol Inform. 2018;9:38. Published 2018 Nov 14. doi:10.4103/jpi.jpi_53_18.
65. Jiang Y., Yang M., Wang S., et al. “Emerging role of deep learning-based artificial intelligence in tumor pathology.” Cancer Commun (Lond). 2020;40(4):154-166. doi:10.1002/cac2.12012.
66. Samek W., Wiegand T., Müller K‐R. “Explainable artificial intelligence: understanding, visualizing and interpreting deep learning models.” arXiv Preprint. arXiv:170808296 2017.
67. Regitnig P., Müller H., Holzinger A. (2020) Expectations of Artificial Intelligence for Pathology. In: Holzinger A., Goebel R., Mengel M., Müller H. (eds) Artificial Intelligence and Machine Learning for Digital Pathology. Lecture Notes in Computer Science, vol 12090. Springer, Cham. https://doi.org/10.1007/978-3-030-50402-1_1.
68. Mukhopadhyay S., Feldman M., Abels E. et al. “Whole slide imaging versus microscopy for primary diagnosis in surgical pathology: a multicenter blinded randomized noninferiority study of 1992 cases.” Am. J. Surg. Pathol., 42 (2018), pp. 39-52. doi: 10.1097/PAS.0000000000000948.
69. U.S. Food and Drug Administration, 510(k) Substantial Equivalence Determination Decision Memorandum – K190332, (n.d.). https://www.accessdata.fda.gov/cdrh_docs/reviews/K190332.pdf.
70. Steiner D., Chen P., Mermel C. “Closing the translation gap: AI applications in digital pathology.” Biochimica et Biophysica Acta (BBA) – Reviews on Cancer, Volume 1875, Issue 1, 2021, 188452, ISSN 0304-419X.