Text as Illumination: Spatial Contrastive Retinex Learning for Language-guided Medical Image Segmentation

A team of researchers has proposed a new framework for language-guided medical image segmentation that treats text descriptions as a form of semantic illumination to improve the delineation of anatomical structures and lesions [1]. The model, called the Text-as-Illumination Retinex Network (TIRNet), was detailed in a paper submitted to the arXiv preprint server on June 26, 2026, by Pingping Zhang and colleagues [1][2]. The work addresses a persistent challenge in Language-guided Medical Image Segmentation (LMIS), where existing systems often lack the fine-grained constraints needed to ensure that the textual prompt and the resulting image segmentation are semantically consistent [2]. TIRNet is inspired by Retinex theory, a concept originally developed to explain human color perception by separating an image into illumination and reflectance components. In this new application, text embeddings function as the illumination source, modulating visual features to highlight relevant regions and suppress irrelevant ones [2]. The architecture integrates two novel components at each decoder stage: a Retinex-inspired Text Modulation Block (RTMB) and a Consistent Detail Compensation Block (CDCB). The RTMB generates positive and negative illumination maps to enhance text-relevant foreground features and suppress background interference, while the CDCB selectively recovers high-frequency details using a consistency-gated mechanism conditioned on the reliability of the illumination [2]. To train the network, the authors introduced a Multi-Scale Illumination Supervision Loss (MSIS-Loss). This loss function combines a Region-Grounded Contrastive Loss, which concentrates cross-modal similarity in foreground regions, and a Background Suppression Loss, which provides pixel-level supervision for the negative illumination maps [2]. The paper reports that TIRNet achieved state-of-the-art performance on the MosMedData+ and QaTa-COV19 datasets [1][2]. The submission file size was listed as 1,127 KB [1]. The research appears on arXiv, an open-access repository for electronic preprints that, as of late 2024, receives about 24,000 new articles per month and is not peer-reviewed [6]. The code for TIRNet has been made publicly available on GitHub [2].