Abstract
Breast implant segmentation on MRI localizer images underpins scanner-side workflows, yet implant cases are demographically rare and exhibit substantial cross-cohort anatomical variation arising from differences in BMI, positioning, implant size, and placement. Deep learning models require adequate representation of such variability to generalize reliably especially challenging in low resolution localizer images. Existing generative augmentation approaches (GANs and diffusion models) can synthesize rare-class samples but often require expert curation, lack paired labels, and must be retrained under distribution shift. We propose a shape-transfer augmentation framework that decouples shape from intensity and texture: breast geometry is transferred from a non-implant pool via constrained deformable registration with continuous, plausibility-constrained (near-diffeomorphic) warps, yielding deterministic, label-paired augmentations. We demonstrate that targeted shape augmentation, when paired with an architecture capable of exploiting it, yields greater gains for implant segmentation than model capacity alone. A <200K-parameter UNet+SAE achieves 0.866 Dice on in-distribution data and 0.872 Dice cross-distribution data (Unseen site data), demonstrating robustness to cross-cohort variation, while outperforming a ∼68× larger nnUNet under distribution shift (0.872 vs. 0.786 Dice).
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/Deep_Brea3th_026.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=X5baEZq8zV
BibTex
@InProceedings{ThoAlp_Shapeshifter_MICCAISAT2026,
author = { Thottupattu, Alphin J. AND Singhal, Vanika AND Bhushan, Chitresh AND Sharma, Pakhi AND Shanbhag, Dattesh D. AND Guidon, Arnaud AND Shriram, K. S.},
title = { { Shape-shifter: Generalized segmentation of breast implants on Localizer MRI } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17256},
month = {pending},
page = {pending}
}