Advancements in Radiomics for Oncologic Imaging: The Role of Deep Learning and AI in Mesothelioma Prognosis

A new study published in European Radiology found that the ResNet-3D-18 model distinguishes malignant pleural mesothelioma from metastatic pleural disease on CT, delivering high diagnostic accuracy (mean AUC 0.972, 95% CI 0.947–0.990) and providing downstream prognostic signal.

The model also supports survival prediction via ensemble classifiers using deep-learning features—an approach that may sharpen noninvasive diagnostic confidence and risk stratification for pleural lesions.

The researchers analyzed pretreatment CTs from 375 subjects (85 malignant pleural mesothelioma, 290 metastatic pleural disease) with lesion- and patient-level evaluation. The team trained a 3D convolutional ResNet backbone on annotated scans with chronological separation into training and independent test cohorts; the primary diagnostic endpoint was lesion- or patient-level classification against a histopathologic reference standard. Training included 70 histologically confirmed mesothelioma and 258 metastatic cases, with an independent test set of 15 mesothelioma and 32 metastatic cases.

For prognostic assessment, deep features informed an ensemble approach. A random forest classifier achieved an AUC of 0.829 (95% CI 0.663–0.943) using 5-fold cross-validation in the 85-patient MPM cohort. The analysis incorporated cross-validated feature selection and internal calibration to estimate discrimination.