Abstract
Breast ultrasound lesion segmentation is challenging due to speckle noise, acoustic artifacts, and high variability in lesion appearance. Although deep learning methods achieve strong performance on benchmark datasets, evaluation is often dominated by overlap-based metrics that may not fully reflect clinically relevant segmentation quality.
In this work, we study the effect of training strategy on segmentation performance and reliability under progressive speckle corruption. We compare standard supervised training with fixed augmentation against a curriculum that gradually introduces spatially correlated multiplicative speckle noise during optimization. Experiments are conducted on the PRECISE/PACE breast ultrasound dataset using U-Net, Attention U-Net, TransUNet, and SegFormer-B0. Evaluation includes overlap-based metrics, boundary-sensitive measures, lesion-level analysis, uncertainty estimation with Monte Carlo dropout, and calibration assessment. Progressive speckle corruption improves Dice across architectures, with the best performance achieved by Attention U-Net (0.8099). However, it reduces IoU and Surface Dice and generally increases HD95. It also leads to higher failure rates and over-segmentation, particularly in malignant and normal cases. In contrast, standard training yields better boundary accuracy, improved calibration, and more reliable uncertainty estimates.
The findings highlight the need for multi-metric evaluation beyond overlap measures when developing segmentation models for artifact-prone ultrasound imaging.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MIRASOL_015.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=teFMHrQeq4
BibTex
@InProceedings{OsmUmm_Beyond_MICCAISAT2026,
author = { Osman-Watara, Ummulkhair AND Afiakwa, Robert Owusu AND Ababio, Frederick Wiafe AND Abdulsalam, Nafisat AND Appiah, Benjamin AND Takal, John AND Bonsu, Felix Osei AND Musah, Toufiq AND Iorumbur, Aondona Moses AND Raymond, Confidence AND Anazodo, Udunna C. AND Ayivor, Livingstone Eli},
title = { { Beyond Dice: Evaluating Reliability and Robustness in Breast Ultrasound Lesion Segmentation Under Progressive Speckle Corruption } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17265},
month = {pending},
page = {pending}
}