Interpretable Deep Learning Models With Concept Whitening Layers

Deep Learning models, such as convolutional neural networks (CNN), are hard to interpret due to their complex, nonlinear, and high-dimensional algorithms. We focused on interpreting CNNs for predicting Hepatocellular Carcinoma (HCC), the most common primary liver cancer, using histopathology images. We used a 50-layer residual neural network (RNN50), an inception CNN, and a recent innovation called concept whitening (CW). The results showed that adding CW layers to CNN models did not significantly affect model accuracy. However, post-hoc analysis methods, including inter-concept similarity rating, intra-concept similarity rating, concept importance rating, and feature vector displays, revealed improved interpretability.