TISC: AI segmentation of TMJ discs achieves 4.96 Dice gain on 1,300 patients

Segmenting the temporomandibular joint (TMJ) disc from MRI is critical for diagnosing internal derangement, but its small size, low contrast, and morphological variability make existing segmentation methods unreliable—often producing fragmented or anatomically inconsistent masks. To solve this, a team of researchers from Korea (including Dayun Ju, Chanyoung Kim, and Seong Jae Hwang) developed TISC (TMJ disc segmentation via semantic anchoring and clinical priors). The framework first uses a Prototypical Semantic Anchoring (PSA) module that aggregates adjacent-slice MedDINOv3 features to create a prototype-driven similarity map, robustly localizing the disc in foundation model feature space. Then, a Clinical-Metadata Point Refinement (C-MPR) module refines boundaries using point-wise predictions modulated by Mouth Open Limitation (MOL), a clinical indicator of disc displacement without reduction.

On a large-scale dataset of 2,488 PD MRI volumes from 1,300 patients, TISC achieved up to a 4.96 Dice improvement over strong baselines across diverse architectures. The resulting segmentations are more anatomically coherent and clinically reliable, reducing measurement instability for disc position and shape. This work addresses a key bottleneck in TMJ disorder diagnosis and demonstrates how combining foundation model features with clinical priors can improve medical image segmentation. The paper is available on arXiv (2606.21177) and is under consideration for publication.