Healthcare · Early-Stage Breast-Cancer Nodule Detection

World Health Organization (Freelance Research Grant)

Role: ML Research Lead (Freelance, Research Grant)MedTech · Computer Vision · Oncology ResearchPrior experience — Michele Sanna, research grantStatus: Completed

Multi-view 3D CNN (adapted MedNeXt) over 420,000 mammograms and 38,000 DBT volumes across 11 countries, with a temporal "prior-exam" branch detecting subtle inter-exam changes. DINOv2-style self-supervised pre-training; federated training across hospitals that could not share raw imagery.

  • PyTorch
  • MONAI
  • DICOM/NIfTI
  • cuCIM
  • Weights & Biases
  • Flower (federated)
0.94 AUCdetection on held-out multi-centre test set
41%of lesions later diagnosed at 12–24 months surfaced on prior-negative exams
−28%false positives at matched sensitivity vs. previous best open benchmark (Mirai-class)
WHO guidanceinformed peer-reviewed publication on AI-assisted screening in LMIC contexts

Next project

Italian RSA Network (Freelance)

Next

Want this kind of work in your business?

Start with the problem, not the platform. Tell us what's slow, manual or breaking.

Send a project briefOr just say hello