Most experimental, programming, writing efforts by Duy Le '27
Abstract: Breast ultrasound studies are limited by scarce pathologist-confirmed
tumour images and variability across scanners. We introduce Tiger-SIREN, an anatomy-aware
synthesis method that repurposes a thyroid-trained U-Net front end with batch-norm
adaptation and a SIREN decoder to graft realistic lesions onto healthy breast scans.
Conditioning is tissue-context rather than text: a soft five-class mask (skin, fat,
glandular, muscle, retromammary) and a coarse box prompt constrain lesion placement
and preserve speckle. On healthy images from the BUS-UCLM cohort, Tiger-SIREN produces
anatomically confined edits with high structural fidelity (measured by SSIM against the
input) and tumour-to-background intensity ratios that fall within the clinical 0.5–0.7
band for most samples, while generating at ∼1 s per image on a single GPU. Because the
pipeline requires only healthy frames and no tumour labels, it offers a practical way
to bootstrap data for breast-ultrasound augmentation. Future work will (i) add automatic
BI-RADS prompt extraction and (ii) quantify downstream impact by fine-tuning a classifier
on BUSI with and without Tiger-SIREN images.
Paper (pdf, preprint)