Advancements in AI for Radiology: Multi-Vendor Validation of Deep Learning in Parotid Tumor Diagnosis

A study found that combined deep learning and radiomics improves preoperative parotid tumor assessment on T2-weighted MRI, with robust automated parotid tumor segmentation and preserved diagnostic discrimination across vendors.

The multicenter, multi-vendor study evaluated 493 pathologically confirmed cases across scanner-stratified cohorts (training n=288, internal validation n=123, external testing n=82) on T2-weighted MRI and reported both segmentation and classification metrics. Automated segmentation achieved DSCs of 0.93 (training), 0.91 (validation), and 0.84 (external testing). Classification AUCs for the integrated deep learning and radiomics model were 0.92 in validation and 0.90 on external testing. The pipeline fused an nnU-NetV2-based segmenter, a ResNet18 image classifier, and a radiomics classifier into a single decision model that retained diagnostic discrimination across cohorts.

Vendor- and sequence-related effects were modest: the largest measurable change was the expected DSC decline on the heterogeneous external cohort, but that drop did not produce a clinically prohibitive loss in diagnostic AUC for the combined model. These findings suggest practical generalizability across common MRI vendors while underscoring the value of local calibration where acquisition protocols diverge substantially.

Limitations merit balanced interpretation: the external DSC decline and imaging heterogeneity emphasize the need for prospective, multicenter roll‑out, and the cohort composition (396 benign, 97 malignant) indicates prevalence-driven effects that should inform local calibration. The highest‑priority next step is a prospective implementation study with real‑world monitoring and formal regulatory assessment as required to confirm clinical utility within routine workflows.