Artificial intelligence (AI) involves the use of machines and algorithms that have the ability to “think” or “act” as if they possessed human intelligence and critical thinking. Machine learning is a type of AI entailing computer software and algorithms that can learn from new data. “Trained” machine learning algorithms can make predictions, execute tasks, make data-driven decisions, and solve complex problems. AI, and machine learning, in particular, have revolutionized, and will continue to revolutionize, a plethora of sectors and industries where powerful data-driven problem-solving abilities are required—the health care industry is no exception!
In medicine, AI and machine learning can be used to guide clinical decision-making, increase the automatization and accuracy of surgical procedures, accelerate drug development, and facilitate health monitoring and disease diagnosis.1 More recently, the applications of AI in digital pathology and biopsy-based diagnosis have also become increasingly evident.
Technology: Understanding AI-assisted digital pathology
In AI-assisted digital pathology, complex algorithms are used to create models that identify key information in digitized whole slide images and make predictions based on this information. As a first step, well-annotated and computer-readable imaging data—also known as structured data—are fed into the algorithm to facilitate algorithm training. These training image datasets are used by the algorithm to generate the best model to maximize the performance and ensure the accuracy of image analysis. Test image datasets of known diagnosis (e.g., cancer tissues and non-malignant tissues) are used to evaluate the diagnostic performance of the algorithms. The patterns formed in the model during the training process are used to make diagnoses and to make decisions, which are then fine-tuned as new data becomes available.2
How can AI and machine learning help advance routine pathology?
Digital pathology generates an immense amount of imaging data, and machine learning is ideal for analyzing the large datasets and the enormous amounts of information derived from digitized specimen slides. Particularly, AI can be used to collect, organize, process, and annotate whole image slides, as well as extract meaningful information from digital pathology images while minimizing the input required from the pathologist.
By integrating AI and machine learning to automate the analysis of digital pathology images, routine diagnosis can become faster, more accurate, and less prone to human errors.3 For instance, Hou et al. demonstrated that the use of convolutional neural networks— a type of deep learning model—to analyze whole slide tissue images could accurately distinguish subtypes of non–small cell lung carcinoma.4,5
Machine learning can also be used to train computers to identify objects in images. The diagnostic accuracy and efficiency of AI algorithms have been found to be comparable to or even exceed those of pathologists.6 In addition to facilitating disease diagnosis and improving the pathologist’s workflow, the ability of AI algorithms to identify objects in images and to extract information on tissue features can be particularly useful for predicting response to different treatments.7
Immunotherapy using immune checkpoint inhibitors has radically transformed the management of various solid malignancies; however, a significant portion of patients do not respond to the treatment. AI-assisted digital pathology can classify tumor tissues based on the extent of immune cell infiltration into highly infiltrated (“hot” tumors), moderately infiltrated, and poorly infiltrated (“cold” tumors).8
Given the strong ability of immune cell infiltrates in the tumor microenvironment to predict response to immunotherapy, AI-assisted digital pathology technologies can be used to distinguish patients likely to respond to immune checkpoint inhibitors from those who are unlikely to respond. By guiding clinical decision-making based on biopsy tissue samples, AI-assisted digital pathology can maximize treatment outcomes and minimize the cost and unnecessary toxicities associated with the use of immune checkpoint inhibitors in non-responders.
Examples of AI algorithms for digital pathology
In a pioneering cross-sectional study evaluating the diagnostic performance of 32 deep learning algorithms developed as part of the researcher challenge competition CAMELYON16, seven algorithms (HMS-MIT II, HMS-MGH I, HMS-MGH II, HMS-MGH III, CULab I, CULab III, ExB I) exhibited higher diagnostic performance than 11 pathologists, providing an area under the curve of 0.994 (best algorithm) versus 0.884 (best pathologist).6
In a 2019 study, researchers from Google AI Healthcare and the Naval Medical Center evaluated the diagnostic performance of their machine learning algorithm Lymph Node Assistant (LYNA). The researchers used LYNA to analyze two datasets (a total of 237 hematoxylin-eosin–stained whole slide images) of lymph node biopsy slides to detect metastatic breast cancer cells. The algorithm accurately distinguished lymph nodes with metastatic lesions from those without metastases 99% of the time. Notably, the algorithm accurately detected suspicious lesions non-detectable by the human eye in a fraction of the time. Furthermore, common histology artifacts, such as overfixation and poor staining, were found not to affect the performance of LYNA.9
The Denmark-based companyVisiopharm offers AI-aided digital pathology solutions to support routine diagnosis and enhance the productivity of pathologists. Their AI-assisted platforms are approved for diagnostic use in Europe; they can be used for breast and colon tissues, among other tissue types, and have been validated for tissue slides stained with H&E or samples stained for Ki-67, ER, or HER2.
Immunoscore (HalioDx) is a commercially available AI-assisted digital pathology platform approved for use with lung cancer and colon cancer specimens to predict the risk of relapse or response to immune checkpoint inhibitors by enumerating peritumoral and intratumoral CD3+ and CD8+ T cells in tumor tissue sections.
Perspectives
With the approval and clinical adoption of digital pathology systems to handle and analyze whole slide images, AI-aided image analysis algorithms are becoming increasingly useful and necessary. As digital pathology generates enormous amounts of data, “smart algorithms” are warranted to handle and analyze these data sets in an automated and robust manner. Therefore, AI can be seen as a key component in facilitating the clinical adoption of digital pathology technologies.
The clinical use of AI-assisted digital pathology technologies in routine diagnosis is expected to increase the accuracy and speed of diagnosis, improve patient stratification, and facilitate the prediction of response to different treatments. Yet, several challenges need to be addressed before we are able to harness the full potential of AI-assisted platforms to analyze digital pathology images for routine diagnosis.
Regulatory issues are the most significant challenge limiting the use of computer-aided diagnostic platforms in clinical practice, as no universally accepted regulatory guidelines currently exist. The inability of regulatory bodies, clinicians and patients, to fully grasp the algorithms and the complex mathematical models behind AI-aided digital pathology platforms, hinders their regulatory approval and widespread clinical use.10 Future clinical studies addressing the diagnostic performance, safety, and error rates ofthese algorithms are required to confirm the feasibility and benefits of using AI-aided digital pathology in routine diagnosis.
Moreover, AI-assisted digital pathology is viewed with skepticism by many pathologists, largely because traditional pathology methodologies have changed little over the years. Training programs to familiarize pathologists with different AI-assisted digital pathology systems and their potentials may help mitigate the mistrust and misconceptions toward these novel technologies.
In addition to not being fully accustomed to computer-aided image analysis systems, medical professionals may fear that the use of intelligent machines may affect job opportunities. The increase in the popularity of AI-assisted technologies raises concerns among medical professionals, who may feel that their jobs are in jeopardy.
Addressing these challenges is paramount, as the clinical implementation of AI-aided digital pathology technologies is expected to revolutionize modern diagnosis and precision medicine by increasing the accuracy of diagnosis and supporting data-driven clinical decision-making.
References
1. Amisha, Malik P, Pathania M, Rathaur VK. Overview of artificial intelligence in medicine. J Fam Med Prim care. 2019;8(7):2328-2331. doi:10.4103/jfmpc.jfmpc_440_19
2. Colling R, Pitman H, Oien K, et al. Artificial intelligence in digital pathology: a roadmap to routine use in clinical practice. J Pathol. 2019;249(2):143-150. doi:10.1002/path.5310
3. Choudhury A, Perumalla S. Detecting breast cancer using artificial intelligence: Convolutional neural network. Technol Heal Care. 2021;29:33-43. doi:10.3233/THC-202226
4. Hou L, Samaras D, Kurc TM, Gao Y, Davis JE, Saltz JH. Patch-Based Convolutional Neural Network for Whole Slide Tissue Image Classification. Proc IEEE Comput Soc Conf Comput Vis Pattern Recognit. 2016;2016-December:2424-2433. doi:10.1109/CVPR.2016.266
5. Sakamoto T, Furukawa T, Lami K, et al. A narrative review of digital pathology and artificial intelligence: Focusing on lung cancer. Transl Lung Cancer Res. 2020;9(5):2255-2276. doi:10.21037/tlcr-20-591
6. Bejnordi BE, Veta M, Van Diest PJ, et al. Diagnostic assessment of deep learning algorithms for detection of lymph node metastases in women with breast cancer. JAMA – J Am Med Assoc. 2017;318(22):2199-2210. doi:10.1001/jama.2017.14585
7. Parwani A V. Next generation diagnostic pathology: Use of digital pathology and artificial intelligence tools to augment a pathological diagnosis. Diagn Pathol. 2019;14(1):19-21. doi:10.1186/s13000-019-0921-2
8. Acs B, Rantalainen M, Hartman J. Artificial intelligence as the next step towards precision pathology. J Intern Med. 2020;288(1):62-81. doi:10.1111/joim.13030
9. Liu Y, Kohlberger T, Norouzi M, et al. Artificial intelligence–based breast cancer nodal metastasis detection insights into the black box for pathologists. Arch Pathol Lab Med. 2019;143(7):859-868. doi:10.5858/arpa.2018-0147-OA
10. Steiner DF, Chen PHC, Mermel CH. Closing the translation gap: AI applications in digital pathology. Biochim Biophys Acta – Rev Cancer. 2021;1875(1):188452. doi:10.1016/j.bbcan.2020.188452
