Eight peer-reviewed papers, mostly oncology: melanoma prognosis, molecular subtyping from whole slide images, and the interpretability methods needed before any of it goes near a clinician.
Journal papers
2025 · Nature Communications
A prognostic model for localised melanoma combining digital biomarkers from the tumour microenvironment with clinicopathological features, across more than 3,500 histology slides from six international cohorts. I built the machine learning and deep learning models and ran the statistical analysis.
It reached a C-index of 0.82 against 0.66 for AJCC staging. The microenvironment features barely moved patients between risk groups. What they improved was calibration, which matters more than it sounds when you're counselling a patient.
2024 · Scientific Data
A dataset of 18,946 dermoscopic images collected at Hospital Clínic de Barcelona between 2010 and 2016, released with baseline classifiers. Built to reflect what clinicians actually see, including nail and mucosal lesions and lesions too big for the dermatoscope aperture, rather than the curated cases that dominate training data.
2024 · Expert Systems with Applications · Open source
A Python package implementing SurvLIME with a matrix-wise formulation for speed, supporting Cox PH, DeepHit and DeepSurv, with visualisation tools for reading the output.
2025 · Computers in Biology and Medicine
A comparison of survival models on a Catalonian cutaneous melanoma cohort, with the salient features extracted and checked against clinical staging.
The features the models leaned on lined up with clinical stage, which is the evidence you need before anyone will let a survival model near a prognosis.
Conference papers
2024 · CVPR Workshop · DEF-AI-MIA
Fine-tuned DINOv2 models with domain adaptation to work around limited pathology data.
BRAF detection worked. Sentinel lymph node prediction did not, because the histopathological correlate is indirect. Reporting that clearly was more useful than reporting a marginal number.
2022 · Springer LNCS
Predicting metastasis from whole slide images of primary melanoma tumours, pairing self-supervised contrastive feature extraction with attention-based multiple instance learning. Attention maps put weight on tumour tissue rather than artifacts.
2025 · Springer LNCS
An attention-challenging MIL framework with a transformer aggregator, strongest with foundation features at lower resolution and larger patch sizes.
2024 · Springer CCIS
Random Survival Forest, DeepSurv and DeepHit against Cox models on SEER data, with feature importance read through SurvLIMEpy.
Take a coffee & chat with me
[email protected]Tell me what you have, whether that's data, video, or a model that already misbehaves, and what decision it needs to support. I'll tell you honestly whether it's tractable.
← Back to the work · Galicia, Spain