Abstract
Reliable tuberculosis screening from chest radiographs requires models that remain robust across acquisition modalities and clinical populations. We present RADAR, a read-out framework that uses a frozen chest-radiograph foundation model and learns only lightweight task-specific heads. RADAR extracts patch representations from multiple depths of the frozen encoder and combines them with a learned tuberculosis query through cross-attention, allowing the classifier to selectively aggregate spatially localized disease evidence. To reduce reliance on acquisition-specific shortcuts, we train complementary heads with modality-adversarial and conditional adversarial objectives, while image-space style randomisation and prediction consistency further encourage acquisition-invariant representations. The resulting heads are ensembled across seeds and input views, with a higher-resolution read-out providing additional sensitivity to small focal abnormalities. The complete system requires no backbone finetuning and uses no acquisition metadata at inference. On MICCAI 2026 TREAT-MMTB Task 2, RADAR achieves an internal F1 of 0.9840 and an external F1 of 0.8375. Across five held-out public sites, the acquisition-adversarial design improves the worst-site F1 from 0.8815 to 0.9064. These results show that a frozen foundation representation combined with an acquisition-aware, attention-based read-out can provide a robust and efficient approach to cross-domain tuberculosis screening.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/TREAT_MMTB_012.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=5qyZJRpe41
BibTex
@InProceedings{RegAth_RADAR_MICCAISAT2026,
author = { Rege, Atharva Atul AND Dukre, Adinath Madhavrao AND Shah, Sarth Santosh AND Razzak, Imran},
title = { { RADAR: Acquisition-Adversarial Attention Pooling over a Frozen Chest-Radiograph Foundation Model for Tuberculosis Screening } },
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}
}