Medical AI
I build deep learning methods for ultrasound images that treat ambiguity — at lesion boundaries and across datasets — as a signal rather than noise. The aim is models that stay reliable on the small, heterogeneous datasets typical of point-of-care imaging, remain small enough to deploy, and show what they based a decision on. Most of this work so far is on breast ultrasound.
01
Loss functions that treat boundary ambiguity as signal rather than noise. The fuzzy rough set loss combines fuzzy similarity with lower and upper approximations, improving Dice by 2.1% on average over the best baseline across breast ultrasound, polyp, brain MRI, chest CT and skin-lesion datasets (CMIG, 2026). An adaptive ensemble loss rebalances BCE, Dice, Hausdorff and Tversky terms during training (MBEC, 2025).
02
Lightweight models that show where they look. Saliency-guided AttentionNet models lesion and surrounding tissue in separate branches: 90.51% accuracy across five public breast ultrasound datasets, and 78.46% on held-out data (BSPC, 2026). MSCT-Trans turns multi-scale CNN features into transformer tokens for breast, thyroid and fetal ultrasound, with Grad-CAM++ localisation and confidence analysis (UMB, 2026).
03
Current work on pixel-level uncertainty estimation, aimed at flagging unreliable segmentations so that only uncertain cases go to human review.
Doctoral work
Advances in Deep Learning for Medical Image Segmentation and Classification — EfficientU-Net, UMA-Net, fuzzy rough set losses, GA-based ensembles and SGAN, validated across five public breast ultrasound datasets.