Abstract
We describe our submission to the Cavity Detection and Segmentation task (Task 1) of the TREAT-MMTB-2026 challenge, which targets binary pulmonary-cavity segmentation on chest X-ray (CXR) evaluated by a composite of case-level detection (0.7) and region Dice (0.3). Detection carries most of the score, yet only 444 mask-labeled images are available. We therefore train the case-level decision on a much larger, publicly available set of case-labeled CXR and keep it separate at inference: a cavity-presence classifier determines the case label, and the segmentation model produces a mask for predicted positives and an empty mask otherwise. Both encoders are adapted to CXR by vision–language pretraining (VLP) on MIMIC-CXR, whose radiology reports we enrich with cavity- and tuberculosis-related sentences from LLM labeling. Rather than using the pretrained encoder directly for segmentation, we first fine-tune it as a cavity-presence classifier on 13,174 case-labeled images, so that it has learned to detect cavities at scale before ever seeing a mask. Pairing a single unaugmented detector with a four-member segmentation ensemble trained on the 444 masks alone scores a composite of 0.6265 on the previously released test split and 0.6001 on the organizers’ held-out external evaluation. Between the two sets, detection transferred well (0.757 → 0.811) while Dice score decreased (0.323 → 0.108). Our model placed third in Task 1.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/TREAT_MMTB_004.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=VMeKa1iJwV
BibTex
@InProceedings{LeeSeo_DetectionGated_MICCAISAT2026,
author = { Lee, Seongeun AND Yun, Hannah AND Jeong, Taejin AND Jeon, Mingyeong AND Park, Junhyun AND Kim, Hyunwoong AND Park, Jonggwon},
title = { { Detection-Gated Cavity Segmentation on Chest X-rays with Vision–Language Pretraining } },
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}
}