Comparative Analysis of CNN Architectures for Automated Knee Segmentation in Medical Imaging: a Performance Evaluation

Abstract

Segmentation of anatomical components is a major step in creating accurate and realistic 3D models of the human body, which are used in many clinical applications, including orthopedics. Recently, many deep learning approaches have been proposed to solve the problem of manual segmentation, that is time-consuming and operator-dependent. In the present study, SegResNet has been adapted from other domains, such as brain tumor, to segment knee bones from Magnetic Resonance images. This algorithm has been compared to the well-known U-Net in terms of evaluation metrics, such as Dice Similarity Coefficient and Hausdorff Distance. In the training phase, various combinations of hyper-parameters, such as epochs and learning rates, have been tested to determine which combination produced the most accurate results. Based on their comparable results, both U-Net and SegResNet performed well in accurately segmenting the femur. Dice Similarity Coefficients of 0.94 and Hausdorff Distances less than or equal to 1 mm indicate that both models are effective at capturing anatomical boundaries in the femur. According to the results of this study, SegResNet is a viable option for automating the creation of 3D femur models. In the future, the performance and applicability of SegResNet in real-world settings will be further validated and tested using a variety of datasets and clinical scenarios.