Binding Text to Pixels: A Narrative Review of Supervision Granularity in Language-Guided Chest Radiograph Segmentation and Grounding
Journal of Imaging · 07.11.2026 · 12 (11) · 559
DOI: 10.3390/jimaging12110559 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Imaging, online erschienen am 07.11.2026. Verfasst von Joseph Kalinzi und 2 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Chest radiographs come with abundant free-text reports but scarce lesion-level annotation, so report text increasingly supervises localization. This structured narrative review, written for model developers, benchmark curators, and evaluators of clinical AI, covers phrase grounding, weakly supervised localization, language-guided segmentation, grounded report generation, and multimodal large language models to September 2026. We propose a five-level granularity ladder, organizing the field by where text binds to the image, not by architecture, as an explanatory framework and testable hypothesis, and show that separating train-time binding from inference-time output exposes the most informative systems. A resource audit classifying each corpus’s spatial signal as reader-drawn, image-derived, or report-derived reveals a structural gap: at MIMIC-CXR scale, lesion masks exist only when generated from the reports that supervise the models, and independent reader-drawn masks for multiple lesion types cover 234–902 images, roughly 400 to 1600 times fewer. On the shared MS-CXR benchmark, three studies report one model at 21.5 to 26.6 mean intersection over union, a spread exceeding most between-method differences. Grounded outputs are intended to make AI assertions checkable at the point of care, though prospective evidence is lacking. We specify an anatomy-referenced evaluation, propose federated subgroup evaluation, and recommend reserving reader-drawn masks for evaluation.
Zitieren
APA
Kalinzi, J., Islam, T., & Ahmed, K. (2026). Binding Text to Pixels: A Narrative Review of Supervision Granularity in Language-Guided Chest Radiograph Segmentation and Grounding. Journal of Imaging, 12(11), 559. https://doi.org/10.3390/jimaging12110559
Vancouver
Kalinzi J, Islam T, Ahmed K. Binding Text to Pixels: A Narrative Review of Supervision Granularity in Language-Guided Chest Radiograph Segmentation and Grounding. Journal of Imaging. 2026;12(11):559. doi:10.3390/jimaging12110559
BibTeX
@article{Kalinzi2026,
author = {Kalinzi, Joseph and Islam, Taminul and Ahmed, Khaled},
title = {Binding Text to Pixels: A Narrative Review of Supervision Granularity in Language-Guided Chest Radiograph Segmentation and Grounding},
journal = {Journal of Imaging},
year = {2026},
volume = {12},
number = {11},
pages = {559},
doi = {10.3390/jimaging12110559}
}
EUVIMED