Review of Deep Learning and Interpretability

Deep learning has revolutionized machine learning. Deep convolution networks play a pivotal role in image classification. A conventional image classification technique requires extensive feature engineering; this approach is exhaustive as rigorous manual efforts are needed. Deep learning automatically detects the feature set required for classification. It uses backpropagation to achieve this. Deep learning is being used in a variety of applications including healthcare. The use of deep learning specifically in cancer biology is significant. For Pathologists, classifying tumors as benign or malignant is a complicated task. Classification of malignant lymphoma is challenging as features are not uniform across whole slides, and there are various patterns for classification of cancer. Deep learning is ideal for such a scenario where automated feature learning is required further deep learning detects features that are invariant to the location in an image. Tumor region-wise classification of a slide can also be achieved which can be further subjected to recurrent neural networks. Textual explanations of cancer regions can be generated, providing further insights into diagnosis. Here, three different cancer datasets are reviewed and subjected to deep learning methods. Explainable AI approaches can be applied to decision trees as well as on gradient boosted trees to obtain explanations.