Deep learning-based severity grading of Meniere's disease using 2D MRI

Meniere's disease (MD) is a complex inner ear disorder characterized by symptoms such as vertigo, hearing loss, tinnitus, and ear fullness, significantly impacting patients' quality of life. Current diagnostic methods lack sufficient accuracy and consistency, highlighting the need for reliable segmentation tools integrated with advanced deep learning techniques.

To develop and evaluate the multi-stage severity assessment system (MSAS), a novel deep learning-based framework designed for precise segmentation and accurate stratification of MD severity using 2D magnetic resonance imaging (MRI).

MSAS was developed using pixel-level manual segmentation on MRI datasets from 189 patients (development cohort) and validated using an independent external test set of 70 patients. The framework integrates sequence-level prediction and slice-level segmentation, utilizing histogram of oriented gradients (HOG), support vector machine (SVM), YOLO-V5, and gradient-weighted class activation mapping (Grad-CAM) techniques. Performance was rigorously evaluated using intersection over union (IoU), dice coefficient, accuracy, mean average precision (mAP), and area under the curve (AUC) metrics.

MSAS demonstrated strong performance, achieving an overall accuracy of 0.971 and an AUC of 0.995. Vestibular region detection yielded mean average precision (mAP) scores of 0.887 (internal set) and 0.877 (external set). For segmentation, dice coefficients were 0.940 (internal) and 0.941 (external), with consistent IoU values of 0.889 across datasets. Enhanced interpretability was achieved through slice attention maps and Grad-CAM visualizations, effectively assisting clinical decision-making.

The results demonstrate the potential of deep learning techniques to enhance MD diagnosis and severity grading, offering significant clinical utility and advancing the understanding of MD's pathophysiology.