February 9, 2023

A Novel Approach to Refining Risk Stratification for Colorectal Patients: Application of Deep Convolutional Neural Networks (DCNN) to Predict Outcome and Molecular Subtyping.

2nd Call, iDoc

Short Summary

With the rising availability of patient data and scans, supervised machine learning has come to a limit. The conventional approaches rely on labeled data that provide feedback to the model on whether or not the predicted outcome is correct. However, to obtain such labels, expert pathologists need to manually annotate large quantities of images which is exhausting and time-consuming. In this project, we aim to use the raw images themselves to learn discriminant tissue features without any kind of manual supervision. Such an approach would naturally be able to differentiate tissue coming from normal to cancerous ones. In addition, the model can be used in complement to clinical scores to help patient stratification and personalized therapy.

Goals

Building a model that is able to properly describe tissue without any manual annotation or prior knowledge. The model can be used for downstream tasks to predict patient overall or disease-free survival.

Significance

Being able to improve daily diagnosis routine. Our model could be used to help pathologists in their decision-making and cancer grading. A better assessment of the cancer subtype is the key to provide to each patient the most adapted treatment to avoid the systematic use of heavy therapies and to improve survival chances.

Background

In 2020, colorectal cancer was 2nd in terms of cancer-related death in Switzerland. Early diagnosis can significantly improve survival outcomes. However, sometimes additional treatment is inevitable. Observations show that patients tend to react differently to certain treatment such as immunotherapy or chemotherapy thus making patient-stratification the key to personalized medicine.

Publications

Abbet, Christian, et al. “Self-Rule to Adapt: Learning Generalized Features from Sparsely-Labeled Data Using Unsupervised Domain Adaptation for Colorectal Cancer Tissue Phenotyping.” Medical Imaging with Deep Learning. 2021.