Enhanced Attention Res-Unet for Segmentation of Knee Bones

Abstract

The objective of this study was to develop a U-net capable of generating highly accurate 3D models of knee bones, in particular the femur. As part of the approach, a U-net was designed, trained, and validated. In order to achieve these goals, a novel architecture was proposed, including an architecture that reduces encoder parameters and incorporates transfer learning, in order to enhance the attention U-net. Additionally, an extra depth layer was added to extract more salient information. Moreover, the model includes a classifier unit to reduce false positives, as well as a Tversky focal loss function, which is an innovative loss function. The proposed architecture achieved a Dice coefficient of 98.05. By using these enhanced tools, clinicians can visualize and analyze knee structures more accurately, improve surgical intervention effectiveness, and improve patient care quality overall.