Automatic segmentation and detection of Vestibular Schwannoma in MRI by Deep Learning

Introduction Automatic segmentation and detection of pathologies in MRI by deep learning is an upcoming topic. We introduce a novel model combining two Convolutional Neural Network (CNN) models for the detection of VS by deep learning.

Methods For deep learning and establishing a combined and new CNN model, our evaluations encompass publicly available (n=242) and in-house datasets (n=96) with high resolution MRI of VS. We used contrast enhanced T1- and T2-weighted MRI slices showing the VS. Images were extracted and annotated with the Software Cascade by an otorhinolaryngologist (ORL) subspecialized in otology and a Radiologist. Experiments were run with 2D, 2.5D, and 3D variations.

Results For the calculation 242 publicly available datasets and 96 in-house datasets of high resolution MRI were used. We annotated 251 T1 and 246 T2 images of the in-house dataset and combined them for the experiments. By the calculation with our new combined CNN Model we were able to show, that the highest Dice Score (0,89) was achieved with T1 images in the 3D Model. In regard to the in-house dataset, the calculation with T1 showed a better detection of VS than with T2.

Conclusion In summary, we present a method enhancing CNN-based VS segmentation performance, validated on public and inhouse datasets, ceT1 and hrT2 modalities to improve accuracy of existing models.

Publication History

Article published online:
19 April 2024