“Deep Learning, Interpretability, and Pathological Cancer Images” 

Speaker: Ehsan Kazemi

Talk summary: Cancer is the most prevalent disease worldwide, and pathological evaluation is pivotal for its diagnosis. Images of tissue sections often vary in appearance through research laboratories. This makes precise diagnosis of cancers’ subtypes and their heterogeneity difficult for a pathologist.  To address this challenge, we investigate the application of deep learning methods to effectively discriminate immunohistochemistry marker staining for different types of cancers.  In this study, we obtain 12140 immunohistochemistry stained whole-slide images of lung, breast, and bladder cancers from Stanford Tissue Microarray (TMA) Database as well as 1509 haematoxylin and eosin stained histopathology images of squamous cell carcinoma and lung adenocarcinoma from The Cancer Genome Atlas (TCGA). Our results suggest that the state-of-the-art deep learning approaches improve the classification accuracy by a considerable margin.  We also try to interpret the predictions of our classifiers by providing qualitative understanding between the tumor images and the response.

Speaker bio: Ehsan Kazemi is a postdoctoral fellow at YINS. He received his PhD from EPF Lausanne. His research is focused on machine learning and data mining. He is also interested in application of these tools in different fields such as social networks and biology.