Improved Alexnet With Transfer Learning Trained on Image Features for Diagnosis of Anterior Cruciate Ligament Injuries

Deep learning aids biomedical image classification, surpassing human experts in some cases. It enhances medical imaging from acquisition to analysis, notably in knee MRI for ACL tear diagnosis, helping prevent joint damage and arthritis. This paper proposes a novel AlexNet with enhanced transfer learning-based anterior cruciate ligament detection (ANet-TL-ACL). The process begins with a preprocessing step to improve the quality and eliminate noise of an image using an adaptive Gaussian filtering (AGF) technique. This enhances the image’s clarity, preparing it for the next stages of analysis. Following preprocessing, the image is segmented using a modified loss and activation function-assisted U-Net (MLA-UNet) model, which precisely identifies and extracts relevant regions, highlighting areas that may contain abnormalities or key features. After segmentation, feature extraction is performed, where distinct features are extracted, such as Pyramid histogram of oriented gradients (PHoGs), morphological features, frequency modulation incorporated local Gabor XOR pattern (FM-LGXP) features, and entropy features. The extracted features are then fed into a detection phase, where an adaptive weighted mixed pooling updated AlexNet (AWM-ANet) model with enhanced transfer learning (ANet-TL) is applied to classify the image. By leveraging pre-trained knowledge, transfer learning boosts the model’s accuracy and efficiency for ACL detection. The performance metrics also show better outcomes while using the proposed ANet-TL technique, when compared with existing methods.