3D Efficient Multi-Task Neural Network for Knee Osteoarthritis Diagnosis Using MRI Scans: Data From the Osteoarthritis Initiative

Abstract

Deep learning, particularly Convolutional Neural Networks, has demonstrated effectiveness in computer-aided diagnosis applications, including knee osteoarthritis analysis. Two of the most common tasks done in medical imaging are the segmentation and classification tasks. This research investigates the feasibility of multi-task models for volumetric analysis using Magnetic Resonance Imaging scans in knee osteoarthritis diagnosis, while considering computational efficiency. In order to leverage the correlation between segmentation and classification tasks, two 3D multi-task models, OA_MTL and RES_MTL are developed to simultaneously segment knee structures and classify knee osteoarthritis incidence. The performance of the multi-task models is evaluated against single-task baseline models and other existing convolutional neural network models using a total of eight different performance metrics, while comparing the computational complexity among the models. Experimental results demonstrate that multi-task model leverages the information of segmentation task to improve the classification performance. OA_MTL is a multi-task model that incorporates an encoder-decoder architecture, residual modules, and depthwise separable convolutions for enhanced performance. OA_MTL achieves superior performance for classification tasks with an accuracy score of 0.825, and a comparable segmentation DSC score of 0.915. OA_MTL achieves a favorable trade-off between computational complexity and model performance. The contribution of this work includes an approach that simultaneously performs knee structure segmentation and osteoarthritis classification in 3D MRI, which addresses the need for efficient models in the field of medical imaging, specifically on computationally challenging 3D medical imaging applications.