Evaluating clinical AI agents in oncology
A framework for judging whether LLM agents that support lung cancer treatment decisions can be trusted, looking at their reasoning as well as their answer. Presented at ESMO 2026.
Clinical AI, medical imaging, neuroscience and NLP. Team projects are marked as such.
A framework for judging whether LLM agents that support lung cancer treatment decisions can be trusted, looking at their reasoning as well as their answer. Presented at ESMO 2026.
Predicted the category of a viewed image from activity in 4,466 visual-cortex voxels, comparing MLP, BiLSTM, Transformer, 3D-CNN and BiGRU models. Handled strong class imbalance with class weights and PCA, and mapped Integrated Gradients attributions back to single voxels.
Fine-tuned ConvNeXt-Tiny on full CT slices and on nodule crops from the LIDC dataset, for benign/malignant and 5-level malignancy. Used MixUp and CutMix against class imbalance and Grad-CAM++ for explanations.
BM25 retrieval and document clustering, fine-tuned BERT and RoBERTa classifiers, prompting of Llama 3.2, and a voice-based retrieval-augmented chatbot.
Fine-tuned EfficientNetV2 to classify 96×96 white blood cell images into 8 classes; trained a U-Net from scratch to segment 128×64 Mars terrain images.
Segmented brain ventricles in 3D multimodal MRI; extracted EEG features to classify sleep stages (REM, NREM, awake).
Multi-label classification of which network edges flood after a simulated storm in Houston, TX, from surface type, water flow, drainage capacity and green-space ratio.