Abstract
Introduction: Many proposed algorithms for tumor detection rely on 2.5/3D convolutional neural networks (CNNs) and the input of segmentations for training. The purpose of this study is therefore to assess the performance of tumor detection on single MRI slices containing vestibular schwannomas (VS) as a computationally inexpensive alternative that does not require the creation of segmentations. Methods: A total of 2992 T1-weighted contrast-enhanced axial slices containing VS from the MRIs of 633 patients were labeled according to tumor location, of which 2538 slices from 539 patients were used for training a CNN (ResNet-34) to classify them according to the side of the tumor as a surrogate for detection and 454 slices from 94 patients were used for internal validation. The model was then externally validated on contrast-enhanced and non-contrast-enhanced slices from a different institution. Categorical accuracy was noted, and the results of the predictions for the validation set are provided with confusion matrices. Results: The model achieved an accuracy of 0.928 (95% CI: 0.869-0.987) on contrast-enhanced slices and 0.795 (95% CI: 0.702-0.888) on non-contrast-enhanced slices from the external validation cohorts. The implementation of Gradient-weighted Class Activation Mapping (Grad-CAM) revealed that the focus of the model was not limited to the contrast-enhancing tumor but to a larger area of the cerebellum and the cerebellopontine angle. Conclusions: Single-slice predictions might constitute a computationally inexpensive alternative to training 2.5/3D-CNNs for certain detection tasks in medical imaging even without the use of segmentations. Head-to-head comparisons between 2D and more sophisticated architectures could help to determine the difference in accuracy, especially for more difficult tasks.