Abstract
Peripheral nerve sheath tumors encompass a heterogeneous group including schwannoma, neurofibroma, and hybrid neurofibroma/schwannoma (HNS). Accurate differentiation is crucial due to distinct biological behavior, different management, and a possible assignment to a genetic disease. We investigated whether deep learning (DL)-based artificial intelligence (AI) reliably distinguished HNS from schwannomas and neurofibromas using whole-slide images (WSIs). Utilizing a dataset of H&E-stained WSIs from 115 tumors, we applied state-of-the-art foundation models (UNI, CONCH, ResNet50) for feature extraction, combined with multiple instance learning algorithms CLAM and mMIL, and a patch-based classifier. Our models achieved high validation performance, with macro-area under the receiver operating characteristic curve (AUC-ROC) values up to 1.0 for UNI-based MIL models. Attention maps correlated well with annotations. HNS were reliably distinguished from schwannoma and neurofibroma using a pretrained AI workflow that was easy to interpret. The choice of foundation model and stain normalization emerged as major performance drivers, with UNI-based mCLAM achieving the highest validation performance, whereas a ResNet50-based mCLAM model with stain normalization reached the best HNS accuracy (0.96) and F1 (0.98) on the test set. Patch-level classifiers with CONCH and UNI embeddings showed slightly lower three-class performance on the validation set (macro AUC-ROC approximately 0.8-0.87), but supported data-driven thresholds for the proportion of Sw-like and Nf-like tissue used for slide-level HNS classification, and a ResNet50-based patch model achieved an HNS accuracy of 0.79 and F1 of 0.88 on the test set. We demonstrated that minimal preprocessing, combined with an interpretable output, quantified tumor components with very high accuracy. This study supports the feasibility of DL tools for the nuanced differentiation of schwannoma and neurofibroma from HNS. Our approach can therefore improve diagnostic accuracy. Limitations include the dataset, technical artifacts, and discrepancies between validation and test performance across encoders. Therefore, future work should include external validation across scanners, labs and centers, protocols, prospective assessment of workflow integration, and benchmarking against interobserver variability to confirm clinical utility and generalizability.