Automatic detection of knee medial collateral ligament (MCL) tear from magnetic resonance imaging using deep neural...

The medial collateral ligament (MCL) is a crucial structure supporting the stability of the knee joint. Although a clinical examination can detect an MCL tear, valgus stress radiography and magnetic resonance imaging (MRI) can confirm it. However, challenges persist in accurate MCL tear detection from MRI images, often leading to misdiagnosis and treatment delays. Therefore, proposing automatic methods for detecting MCL tears is necessary. To the best of the researcher's knowledge, studies have yet to address this problem using deep neural networks.

This study aims to detect medial collateral ligament (MCL) tears through knee MRI images. Our collected dataset includes coronal views of knee MRI images taken from patients in "Imam Khomeini Hospital, Tehran, Iran." The dataset has 3575 knee MRI images from 60 patients, each with a resolution of 512 × 512 pixels. We employ the Cross-Industry Standard Process for Data Mining (CRISP-DM) methodology to structure our approach, which includes three distinct scenarios utilizing deep learning models. In Scenario One, A custom Convolutional Neural Network (CNN) architecture is designed specifically for MCL tear detection. This model undergoes a meticulous fine-tuning process and is evaluated using a comprehensive knee MRI dataset. We adapt the hyperparameters of the CNN to accurate optimize its performance, ensuring classification of MCL tears. In the second scenario, we leverage a deep neural network pre-trained on the ImageNet dataset. The pre-trained VGG19 model is utilized, where we extract features from its layers and feed them into custom output layer to classify MCL tears using fine tuning. This approach allows us to assess the effectiveness of transfer learning in improving MCL tear detection. In the third scenario, we implement transfer learning using the VGG19 architecture. We apply transfer learning by freezing the early layers of the pre-trained VGG19 model and modifying its final layers. This innovative approach aims to enhance the model's ability to accurately identify MCL tears by utilizing learned features from a large dataset.

The experimental results demonstrate that the first scenario achieved an accuracy of 95 %. The second scenario outperformed the others, achieving an average accuracy of 98.3 %, an average loss of 0.07, and an area under the receiver operating characteristic (ROC) curve (AUC) of 1.00. In contrast, the third scenario attained an accuracy of 80 %.

This study highlights the effectiveness of deep learning, particularly the pre-trained VGG19 model, in detecting MCL tears from knee MRI images with 98.3 % accuracy. By leveraging transfer learning, our approach mitigates data limitations, demonstrating the potential of automated diagnostic tools to improve accuracy and efficiency in clinical practice.

Our study introduces a highly accurate deep learning-based method for detecting MCL tears, potentially enhancing timely diagnoses. Integrating this approach into CAD systems could improve patient outcomes by supporting medical decision-making. Future research should validate these findings across diverse populations to ensure robustness.