Research
I work on machine learning for medical imaging and clinical language, with a focus on models that learn from the imperfect supervision real hospitals produce.
Multimodal learning for 3D medical imaging
Radiology produces images and reports together, and I study models that use both. In my PhD, I train end-to-end 3D vision-language models that receive a full CT volume and a free-text finding from its report, and predict the corresponding 3D mask.
Before that, my MSc thesis trained multi-label classifiers over frozen CT-CLIP volumetric features and used linear probing to isolate a representational bottleneck.
Learning from imperfect supervision
Hospitals have far more reports than expert labels. I study how far models can go with the supervision that is actually available: labels generated by language models, partial voxel-level annotation, and long-tailed findings.
In my undergraduate thesis, large-scale weak supervision with GPT significantly outperformed scarce expert labels for annotating Spanish chest CT reports. In my MSc thesis, synthetic supervision matched expert labels from about 100 examples and outperformed them beyond that. My PhD extends this line to positive-unlabeled learning under partial annotation and to measuring how much models rely on text versus image evidence.
Language technologies for clinical text
Much clinical knowledge lives in free text, often in Spanish. I have worked on extracting structured information from mammography reports with large language models and retrieval-augmented generation, on detecting contextual laterality, and on generating mammography reports.
Beyond clinical text, I have contributed to research on multilingual minimal contrastive editing.