THE ROLE OF ARTIFICIAL INTELLIGENCE IN PATHOLOGIC DIAGNOSIS

Abstract

Over the last ten years, there have been thousands of papers describing the potential role of artificial intelligence (AI) algorithms in medical decision making. The impact of AI breakthroughs seems to be major in the field of medical image analysis in which such algos are supposed to do a better job than medical doctors, especially in radiology. Although the use of AI in medicine will increase over time, today the use of AI algos in medical decision assistance is quite limited. Radiology images are directly obtained from a machine whereas histopathology images are converted (digitized) from colored tissue sections on glass slides. The latter represent a tremendous source of heterogeneity explained by inter laboratory variations of tissue section processing. So far, most of the algos proposing automatic analysis of histopathology images are based on convolutional neural networks (CNN) which are extremely sensitive to variations of image heterogeneity and thus prone to overfitting. In other words, an algo trained on pictures produced in lab A is unable to recognize the same pictures obtained from lab B. In this presentation, different alternatives to circumvent CNN (deep learning) limitations will be presented. One of the strategy is to use large scale AI models so called foundation models (FM) based on transformer architectures which are trained on huge amounts of pictures. Another approach, less costly in terms of data input for training, is to use cartesian genetic programming (CGP) algos which, in addition, fill the gap of explainability of decision making. The final goal is to come out with AI solutions proposing a robust (accurate) assistance in picture analysis that can be run in every pathology laboratory regardless of tissue processing variations.

Keywords: bioinformatics; computational and systems biology; pathology and classification of lymphomas

No potential sources of conflict of interest.