Automated Segmentation for Knee Joint MRI Images Using Hybrid UNet+Attention

Abstract

Automated segmentation of knee subchondral bone structures such as area and shape using deep learning approaches is a significant task for medical MRI images. However, existing techniques usually suffer from many challenges due to complex tissue structure when utilized in 3D due to their large memory requirements, and unusual image contrast/ brightness. This paper aims to exhibit proof of the concurrent effectiveness and reliability of the dynamic segmentation technique currently used to quantify 3D statistical shape/image-based in knee assessment and to propose suggestions for their utilization in the treatment of osteoarthritis disease. The proposed automated Hybrid UNet+Attention technique involves the enhancement of contrast of knee MRI bone surface images and can process large full-size 3D input samples (no patches) within seconds using the CPU. The overall performance of the proposed technique was estimated against ground truths by computing performance metrics like Intersection over union (IoU), dice similarity coefficient (DSC), precision, and recall.