Neuroimaging Cranial Pathology Delineation through Contrast-Specialized Specialist Routing
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Purpose Cranial MRI reaches the reading room in whatever contrast combination the scanner, site and protocol produced, and lesion appearance changes with contrast. A department that wants segmentation support must maintain a separate model for each contrast combination, a library that is hard to sustain in routine practice. We evaluated a single deep learning framework that replaces that library with contrast-specialised routing. Materials and methods The framework (OMNIA) keeps one specialist subnetwork for each MR contrast and fuses their predictions at every resolution level through a hierarchical gating network that assigns voxelwise weights. Progressive training protects each specialist's contrast-specific behaviour, and a soft allocation module routes incomplete contrast sets to the specialist channels under an availability prior. The study was retrospective, across nine collections totalling 6,336 annotated volumes, five contrasts and six pathology groups, with a healthy reference collection; six public collections were used for training and internal testing, and three institutional collections were held back for external testing. Results On the seen collections the framework reached a mean Dice of 0.8204, above every universal baseline (0.8007–0.8084) and close to dedicated per-combination pipelines (0.8250), as a single deployable model. On the unseen collections it reached 0.7951, the strongest result among single models. On multi-contrast BraTS 2021 input it scored 0.8334 against 0.8366 for fifteen dedicated models, at about one third of their storage (1.65 GB versus about 4.95 GB). Conclusion Contrast-specialised routing allows one deployable model to approach the accuracy of dedicated per-combination pipelines, on data and contrast combinations beyond those it was trained on.