Research

3D breast ultrasound: telling malignant lesions from benign

Automated 3D breast ultrasound (ABUS) images the whole breast in one sweep, but reading the volumes is slow and separating malignant from benign lesions is hard. My research asks what the shape of a lesion's surface says about it.

1. Thesis. MSc at IUST (2023): classification of breast cancer lesions in 3D-ABUS images, supervised by Dr. Mohsen Soryani and Dr. Ehsan Kozegar.
2. Published method. Laplace–Beltrami spectra of the lesion surface, fed to a dual-path CNN: AUC 0.935, accuracy 84.3% on the official TDSC-ABUS test split (70 lesions), as reported in the paper. Paper (DOI)
3. Current work. Why the spectrum works: the mathematics of Laplace–Beltrami shape representation, and better descriptors derived from it. It starts from a fair benchmark: classical baselines reproduced under one protocol (Tan et al. 2012 reaches AUC 0.86 on the same split). Project page · Code
4. Tooling. tdsc-abus2023-pytorch, the dataset loader everything above runs on: GitHub · PyPI
5. Write-ups. Part 1: mammography, ultrasound, and why 3D

A caveat that matters: the reproduced baselines use the ground-truth lesion masks, while challenge entries had to find and segment lesions themselves (the winner reached AUC 0.889), so those numbers are not directly comparable. Running every method on predicted segmentations is part of the current work.