Lectures and seminars Faculty event: Bojing Liu
Title: Identifying Interpretable Histomorphological Features of Colon Cancer through Self-Supervised Learning-Based Automation
Speaker: Bojing Liu
Self-supervised learning (SSL) in artificial intelligence, requiring no histopathology annotations, efficiently extracts features from hematoxylin and eosin-stained whole slide images (WSIs). SSL enables exploring the potential of histopathology to predict colon cancer overall survival (OS) in an interpretable manner. The Barlow Twins-encoder was trained on 435 TCGA-colon adenocarcinoma WSIs to extract features from small image patches (tiles). From the extracted features, tiles were grouped into distinct histomorphologic phenotype clusters (HPCs), used to predict OS. Pathologists analyzed the histologic tissue composition of each cluster and linked them to immune and gene expression profiles. Feature reproducability and performance of prediction was validated in an unseen cohort from the AVANT-trial (n=1213 WSIs). We derived 47 unique HPCs, which were placed in larger cluster groups through blinded histopathological analysis. Clusters with immune cells, healthy colon tissue, and aligned tumor stroma associated with better OS, while mucinous, poorly-differentiated tumors or disorganized tumor stroma were linked to worse OS. Immune profiles with high leukocyte fractions correlated with immune cell-rich HPCs. In the AVANT-bevacizumab group, clusters linked to better survival showed enrichment in gene pathways promoting VEGFa production, the target for bevacizumab. Overall, OS prediction from our SSL-based histopathology model achieved state-of-art performance and outperformed the clinical baseline model including variables age, sex, tumor-node-metastasis staging, and tumor-stroma ratio. The SSL Barlow Twins-encoder effectively formed and reproduced distinct HPCs of WSIs from well-documented cohorts. Utilizing the clinical AVANT-trial, we validated clusters linked to OS in standard-treated colon cancer patients and those treated with bevacizumab, revealing novel histopathology insights and emphasizing the significance of tissue type quantity and tumor stroma architecture in relation to OS.
About the speaker
I completed my Ph.D at MEB under the supervision of Dr. Karin Wirdefield. I have been a postdoctoral fellow at NYU Langone working with Dr. Aristotelis Tsirigos. Currently, I am a postdoctoral in Dr. Mattias Rantalainen’s research group.
Join the faculty event in Wargentin! After the seminar, faculty lunch will be served in Ljusgården.