Do MRI radiomic models truly generalize? External validation of three studies in parotid lesion characterization

External validation of six radiomic models published in three studies: two distinguishing benign from malignant lesions (study 1) and four distinguishing pleomorphic adenomas from Warthin's tumors (studies 2 and 3).

This monocentric retrospective study included 133 patients who underwent MRI before parotid tumor surgery at our center from 2005 to 2022. For study 1, T1 and T2FS images of 109 benign lesions and 21 malignant ones were included. For study 2, T1 and T2FS images of 58 pleomorphic adenomas and 34 Warthin's tumors were included. For study 3, T2 images of 35 pleomorphic adenomas and 16 Warthin's tumors were included. After segmentation and extraction of the radiomics parameters, the radiomics (Radscore) and combined clinical and radiomics (Nomoscore) models from all 3 studies were applied. Performance was also studied after ComBat harmonization for multiple scanners. Performance was studied on all patients and for studies 1 and 2 on a subgroup of 58 patients who had undergone their examination on the same MRI machine.

AUCs were 0.540/0.548 (Radscore/Nomoscore) for study 1, 0.521/0.521 for study 2, and 0.639/0.630 for study 3, whereas the AUCs in the original studies were 0.908/0.938, 0.902/0.918, and 0.796/0.934, respectively. The results were similar after ComBat harmonization. In the subgroup analysis, the AUCs were 0.533/0.538 for study 1 and 0.513/0.516 for study 2.

Our external validation study was unable to reproduce the results of the six published radiomic models for characterizing parotid lesions, suggesting the limited applicability of these radiomic tools in clinical practice.

Question We aimed to perform an external validation of six previously published MRI radiomic models for the characterization of parotid lesions. Findings The performances on our population of the six radiomic models were lower than in the initial studies, the highest AUC being 0.639. Clinical relevance Our study failed to replicate the performance of the six previously published MRI radiomic models for the characterization of parotid lesions, indicating that the clinical applicability of these radiomic approaches is limited.