Anatomically Consistent TMJ Disc Segmentation via Semantic Anchoring and Clinical Priors

Researchers have developed TISC, a novel AI framework for accurately segmenting temporomandibular joint (TMJ) discs from MRI scans by combining semantic anchoring with clinical metadata. The method achieves up to 4.96 Dice improvement over existing approaches and produces anatomically consistent results for more reliable diagnosis of internal derangement.

Medical image segmentation remains a persistent challenge in diagnostic imaging, particularly for anatomically complex structures like the TMJ disc that exhibit high morphological variability and low contrast in MRI. Traditional segmentation architectures, when applied to this specialized domain, frequently generate fragmented or anatomically implausible masks that compromise downstream clinical measurements. TISC addresses these limitations through a domain-aware approach combining two innovations: Prototypical Semantic Anchoring leverages multi-slice feature aggregation from foundation models to establish robust localization, while Clinical-Metadata Point Refinement incorporates clinical indicators like Mouth Open Limitation to guide boundary refinement with anatomical constraints.

The framework's performance on a large-scale cohort of 2,488 MRI volumes demonstrates the value of embedding clinical priors into deep learning architectures rather than treating medical segmentation as a generic computer vision problem. This hybrid approach—combining foundation model capabilities with domain-specific constraints—reflects a broader trend in medical AI toward more interpretable, clinically-aligned systems that account for the semantic knowledge embedded in clinical practice.

For medical imaging developers and healthcare institutions, this work suggests that substantial improvements in diagnostic reliability require moving beyond architecture innovation alone. The 4.96 Dice improvement represents a meaningful advancement in segmentation quality that directly impacts measurement stability for clinical decision-making. Healthcare providers implementing AI-assisted diagnosis gain access to more reproducible assessments, reducing inter-observer variability and potentially enabling earlier detection of TMJ dysfunction. The methodology's emphasis on anatomical consistency over raw accuracy metrics offers a template for other medical segmentation tasks involving small, variable structures.

• →TISC framework integrates semantic anchoring and clinical metadata to improve TMJ disc segmentation by up to 4.96 Dice points over baseline methods
• →Prototypical Semantic Anchoring module establishes robust localization by aggregating multi-slice foundation model features into similarity maps
• →Clinical-Metadata Point Refinement incorporates Mouth Open Limitation indicators to guide anatomically consistent boundary predictions
• →Evaluated on 2,488 MRI volumes from 1,300 patients, demonstrating scalability and clinical relevance of the approach
• →Method addresses fragmentation and anatomical inconsistency issues plaguing existing general-purpose segmentation architectures in medical imaging