🏥 Overview Master thesis at Fondazione IRCCS Istituto Nazionale dei Tumori di Milano. The goal was 6-month overall survival prediction from lung CT scans in advanced NSCLC patients, with explainability and fairness built in from the start.
Clinical Motivation
Advanced NSCLC patients receiving immunotherapy face highly variable outcomes. A CT-based model that can reliably flag which patients are unlikely to survive 6 months could materially inform treatment planning, but only if that model is interpretable, fair, and evaluated without leakage.
🔬 Technical Scope Developed and evaluated 2D and 3D computer vision pipelines for 6-month overall survival prediction. Trained CNN-based models (including ResNet50) and benchmarked against Vision Transformer and GAN-based approaches. Worked on the Apollo11 cohort: 385 advanced lung cancer patients treated with immunotherapy. Applied SmoothGradCAM++ for spatial explainability, identifying which CT regions drove model predictions. Performed fairness assessments showing that naive augmentation strategies can silently introduce bias, something most papers don’t check for. 📊 Results 0.74 F1 on 6-month overall survival prediction, a hard binary classification task on a real-world clinical cohort (Apollo11, 385 advanced NSCLC patients treated with immunotherapy).
The fairness analysis identified a previously unreported failure mode: naive augmentation strategies can silently introduce demographic bias, a finding with direct implications for how medical imaging models should be evaluated before clinical deployment.
📢 Dissemination 4 peer-reviewed outputs from a single master thesis, spanning AI, clinical oncology, and trustworthy AI.
...
