Interpretable deep learning in healthcare

Description

Artificial intelligence, especially deep learning, has made major breakthroughs in performing various complex tasks such as computer vision, speech recognition, and natural language understanding. Despite the widespread success of these deep learning models, they are usually regarded as “black box” solutions. Often better predictive power comes with deeper models, which makes it harder to understand and interpret why a specific outcome has achieved. But in a sensitive discipline like healthcare, where any decision comes with a huge long-term responsibility, qualitative and quantitative evaluation of how decisions are made is essential for acceptance and integration of AI into the clinical workflow. It is also of utmost importance to assure regulators that appropriate evidence supports these AI-based clinical decisions.

Behrooz Hashemian shares various use cases of artificial intelligence in healthcare that are set to dramatically change the clinical practice. Behrooz outlines the main challenges of integrating AI in hospitals and clinical workflows, including interpretability of deep learning models. Interpretability is not only important for integration but also essential for improving the model by showing why it fails, increasing the security by discovering the edge cases, and verifying the reliability and transferability of the models by reasoning about the decision process. Behrooz reviews interpretability methodologies, from feature visualization to saliency maps and deconvolution, and discusses their power and limitation. He concludes by demonstrating how he and his team have exploited these state-of-the-art methods for characterizing different lesions in medical images, paving the way for adoption and integration of AI in hospitals.