DriftNet: A Contrastive Learning for Pan-Cancer Stage Classification

Abstract

Identifying the stage of cancer is crucial because it greatly helps the chances of successful treatment. In this study, we developed a new intelligent system, named DriftNet, for the early detection of cancer stages. First, large DNA methylation data are compressed into smaller meaningful signals using an autoencoder. Then, we trained the model adopting contrastive learning to better distinguish cancer stages, which pulls similar cases together and pushes apart different ones. Finally, we combined these with simple patient information such as age and cancer type in a two-branch deep learning model. DriftNet is tested on the TCGA dataset, from 1,343 patients across multiple cancer-types. In the final evaluation, DriftNet achieved the highest test accuracy of 67.82%, the highest macro F1-score of 0.6442, and the highest Stage III F1-score of 0.5138. It outperformed both classical machinelearning baselines and paper-inspired deep-learning baselines, especially for the difficult Stage III class. These results show that supervised contrastive learning and dual-branch fusion can improve pan-cancer Stage I/II/III classification using methylation and clinical information.