Abstract
Accurate cancer prognosis is critical for personalized treatment but remains challenging due to high-dimensional, heterogeneous biomedical data and limited sample sizes. This dissertation proposes a novel deep learning framework integrating supervised contrastive learning (CL) and a semi-supervised mean-teacher model, both optimized for survival prediction via the Cox proportional hazards loss.
First, we apply supervised CL to whole-transcriptome gene expression data across 19 cancer types from The Cancer Genome Atlas (TCGA). This approach yields compact, biologically meaningful embeddings that significantly enhance the concordance index (c-index) and risk stratification compared to traditional methods. Notably, the model exhibits robust out-of-distribution generalization on external CPTAC and DKFZ cohorts without retraining.
To improve robustness and leverage unlabeled data, we introduce a semi-supervised mean-teacher architecture. By enforcing student-teacher consistency using the proposed Cox loss, this approach achieves a $>10\%$ increase in c-index over supervised baselines. Incorporating a large unlabeled external BRCA cohort further elevates the TCGA BRCA prognostic c-index to 0.89.
The framework is subsequently extended to multimodal analysis by integrating diagnostic whole-slide images (WSIs). We developed a pipeline to automatically segment tumor regions and encode features using pre-trained models, including ResNet and DINOv2. These WSI features are fused with genomic data via a mutual attention mechanism, yielding an additional 5% improvement in the c-index over the unimodal model.
This dissertation presents a unified semi-supervised, multimodal framework that synergizes gene expression and imaging data. The resulting methods and open-source tools establish a scalable foundation for next-generation precision oncology, effectively bridging advanced deep learning with clinical practice.