Artificial intelligence (AI), a collective term for a wide variety of machine learning systems, has progressed significantly in recent years with the development and widespread dissemination of deep learning techniques. Deep learning is a specific machine learning approach that uses neural node architectures reminiscent of those found in the human cortex. Neural networks can be trained on large quantities of data, allowing them to develop feature recognition capabilities that permit discrimination between various patterns in a data set. Deep learning approaches have been shown to function at a human – or even superhuman – level in various domains, recently beating a world-class player at the highly complex and intuitive game of GO (1).
The implementation of machine learning approaches for medical diagnostics has long been a topic of interest, but translation to real-world settings has remained limited (2). However, with recent developments in deep learning, the possibility of sophisticated decision support for clinicians has been aggressively rejuvenated. A flurry of publications in recent years have demonstrated the potential for deep learning applications in such varied fields as dermatology, ophthalmology, oncology, radiology, and pathology. Radiology and pathology, in particular, are considered highly amenable to deep learning-based technologies, given the particular strengths of these algorithms in image analysis (3).
For pathology, deep learning approaches carry significant potential to improve the diagnostic accuracy and daily workflow efficiency of practicing pathologists. Various groups have examined and demonstrated the inter-observer variability present between pathologists who have assessed a single set of cases (4). The introduction of sophisticated algorithms trained on large quantities of data (previously annotated by qualified pathologists) has the potential to improve the consistency of diagnostic decisions. Artificially intelligent systems can be leveraged to not only provide diagnostic outputs, but also examine submitted data sets to identify correlations between patient prognosis and subtle morphologic variants that humans cannot yet recognize. One can further envision the automation of tedious, repetitive tasks: ordering anticipated stains for a submitted section; quantifying features such as mitotic count or percent positivity in immunohistochemically stained sections; populating final diagnostic reports. The concomitant efforts of human and machine diagnosticians may provide an additional layer of quality assurance, reducing the rates of analytical and post-analytical errors present in pathology departmentstoday (5).
Discussions about AI invariably prompt questions and concerns about the possibility of displacing human pathologists from their roles. Evolution of the relationship between man and machine is extremely difficult to predict, as evidenced by the wide range of opinions on the matter. The question is further complicated by the non-intuitive rate of development of novel technologies. Among practicing pathologists today, most believe that these platforms will eventually play a role in diagnostic pathology – but primarily for decision support, rather than clinician-independent analysis.
Perspectives of computer scientists at times diverge from those of clinicians. In radiology, some leading AI researchers have expressed a more dramatic view, proposing that new modalities will likely displace radiologists to some degree. The potential impact on pathology seems to attract less comment – probably due to the reduced visibility of the field, but possibly because some technical aspects of pathology (such as identification of tissue orientation, or the accurate determination of margins and tumor extent) may be less amenable to automation.
In short, AI brings both promise and challenge. Many questions remain unanswered, but will need to be addressed in time – especially considering that, given recent developments, the integration of artificially intelligent tools seems very likely. In the near future, it will be important to develop the technical and intellectual infrastructure necessary to permit smooth and effective uptake of new technologies. Robust involvement by clinicians in the development and implementation of these tools may permit increased control over the process, increasing the chances of effective and productive integration of new approaches.
Newsletters
Receive the latest pathology news, personalities, education, and career development – weekly to your inbox.
References
- D Silver et al., “Mastering the game of Go without human knowledge”, Nature, 550, 354–359 (2017). PMID: 29052630. I Kononenko, “Machine learning for medical diagnosis: history, state of the art and perspective”, Artif Intell Med, 23, 89–109 (2001). PMID: 11470218. A Madabhushi, G Lee, “Image analysis and machine learning in digital pathology: challenges and opportunities”, Med Image Anal, 33, 170–175 (2016). PMID: 27423409. A Harbias et al., “Implications of observer variation in Gleason scoring of prostate cancer on clinical management: a collaborative audit”, Gulf J Oncolog, 1, 41–45 (2017). PMID: 29019329. SC Hollensead et al., “Errors in pathology and laboratory medicine: consequences and prevention”, J Surg Oncol, 88, 161–181 (2004). PMID: 15562462.