Machine Learning and Veterinary Pathology: Be Not Afraid!

This issue of Veterinary Pathology features a timely review by Awaysheh et al ¹ that provides a historical perspective of machine learning in human and veterinary medicine, an overview of machine learning categories and supervised learning algorithm architypes, and a discussion on input data variables that affect machine learning and predictive outcomes.

During the recent course of strategic planning conducted by the American College of Veterinary Pathologists (ACVP), a landscape assessment informed by member and external stakeholder surveys revealed that computer technology and molecular medicine are significant disruptive forces for the future of veterinary pathology. Yet artificial intelligence and advanced molecular tools are driving the movement toward precision medicine, which tailors treatment to the individual characteristics of patients, both human and animal, based on classification into subpopulations that differ in disease susceptibility, pathogenesis, morphology, response to treatment, and/or prognostic outcomes. ¹⁰ While humans can learn to make general and complex associations based on small amounts of data, the voluminous data sets available through electronic medical records, laboratory information systems, “multi-omics” technologies, in vivo diagnostic imaging, and digital pathology photographs and whole-slide images are overwhelming for a single patient, let alone larger subpopulations. ¹³

The intelligence demonstrated by machines to mimic human decision-making processes may be achieved through supervised, unsupervised, and reinforcement learning approaches to achieve predictive properties based on example data. ³ Such example data are provided in the form of inputs/features and outputs/labels and include audio, video, images, speech, and text. In deep learning, a subset of machine learning, deep neural networks are developed so computers can map information and predict classification of new inputs. ^1,13 Even though computer algorithms can extrapolate patterns and expose correlations that suggest causality, they must be trained and are unable to identify causal links. ⁴ So be not afraid…machine learning will enhance, not extinguish, pathologists! Of utmost importance is the need for pathologists to provide the “ground truth” or specific label, be it a biomarker, diagnosis, or annotation of a feature such as a mitotic figure. ¹³ Once machine learning models are trained, they work in tandem with pathologists to provide improved diagnostic results with increased efficiency and accuracy. ⁵ New clinicopathological relationships may also be revealed, leading to the generation of new hypotheses when unsupervised learning methods, such as clustering algorithms, analyze data without prior knowledge of inputs/features and categorize into distinct cohorts with unique features.

However, the resultant predictive data provided by machines through deep neural networks are only as valuable as the input data they receive. The review herein by Awaysheh et al ¹ of variables that affect machine learning processes is widely applicable to veterinary pathologists, especially those dipping their toes in the machine learning pond. A number of the points discussed can be reinforced upon application to quantitative image analysis of whole-slide images using supervised software, for example. First and foremost, algorithms should be selected or developed based on specific projects and research questions to be answered. It is also important to understand the array of preprocessing steps available to extract valuable information from the pixels representative of the original glass slide. Feature selection should invariably involve a pathologist at some point; pathologists must consider the number of features for machine learning just as they contemplate severity levels when creating or adapting grading schemes, because more is not always better. Performing annotations at lower magnifications (×200 vs ×400) can also have variable effects on the prevalence of false positives or false negatives. ^2,8 In addition, variability in routine hematoxylin and eosin staining as well as immunohistochemical staining can affect machine learning through incorrect classifications. ^6,7 Pathologists are, understandably, more sensitive to manipulations in color, whereas imaging scientists are more sensitive to image noise, which can be introduced at the time of image acquisition or transmission and can undergo preprocessing. ^11,12

The concept of text segmentation preprocessing, especially when investigating clinicopathological relationships, is also an interesting one to consider in veterinary pathology, particularly when electronic medical record and laboratory information systems vary widely between academic veterinary medical centers as well as state and commercial diagnostic laboratories. While the Systematized Nomenclature of Medicine–Clinical Terminology (SNOMED CT) is not consistently applied in all veterinary medicine settings, standardized nomenclature as proposed in the International Harmonization of Nomenclature and Diagnostic (INHAND) Criteria strives to minimize textual variability in preclinical studies mined from databases used in the biopharmaceutical industry. ⁹

What we as pathologists do with the results provided by deep learning models is what positions pathology on the trunk of the tree of medicine! We draw from all basic sciences to dissect mechanisms of disease and integrate findings throughout the entire biologic system—whether zebrafish, rodent, faithful canine companion, dairy cow, racehorse, or cheetah—to discover new approaches to diagnose, treat, and prevent disease.

With the 2018 to 2022 Strategic Plan, ACVP intends to propel veterinary pathology into a future in which our collaborative teams include not only clinicians, pathologists, molecular biologists, and biomedical engineers but also image and computer scientists as well as bioinformaticians. Consequently, it is essential that veterinary pathologists have a basic understanding of machine learning systems and their limitations, as well as incorporate these fundamentals into the training and assessment of our future pathologists. Veterinary Pathology, the official journal of the American, European, and Japanese Colleges of Veterinary Pathologists, bridges the divide between natural and experimental diseases of animals (https://us.sagepub.com/en-us/nam/veterinary-pathology/journal201966#description). As such, manuscripts that increasingly incorporate machine learning in diagnostic investigations of natural and emerging diseases, experimental studies that use animal models and elucidate mechanisms of specific disease processes, and characterizations of environmental and pharmaceutical hazards are encouraged and welcome.