Abstract
Tuberculous cavities are challenging to detect and segment on chest X-rays because they are often small, low contrast, and obscured by overlapping thoracic structures. TREAT-MMTB Task 1 jointly evaluates image-level cavity detection and pixel-level cavity segmentation. We propose a global-to-local framework that integrates automatic cavity localization, local segmentation, and image-level detection using a pretrained CheXFound ViT-L/16 encoder. Full-field probability cues and learned spatial features are used to generate candidate regions, which are refined by a local segmenter adapted to automatically predicted regions. Five fold-specific segmentation maps are restored to the native image grid and averaged before thresholding to obtain the segmentation mask. Image-level cavity presence is estimated from segmentation- and proposal-derived evidence and further refined using complementary whole-image context. In five-fold out-of-fold evaluation on 444 development cases, the proposed framework achieved a classification accuracy of 0.8243 and a Dice score of 0.3225 for cavity segmentation.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/TREAT_MMTB_010.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=jbOAsE3Q8x
BibTex
@InProceedings{LimNay_GlobaltoLocal_MICCAISAT2026,
author = { Lim, Nayeon AND Cha, Sojeong AND Choi, Seoyeon AND Jeong, Juhyeon AND Shin, Taehoon},
title = { { Global-to-Local Segmentation and Context-Aware Detection of Tuberculosis Cavity in Chest X-rays } },
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}
}