Artificial intelligence in medical diagnosis (with emphasis on orthopedics)

Abstract

Early and accurate diagnosis of hip conditions, such as fractures and degenerative diseases, is crucial for ensuring that patients receive appropriate treatment on time. Delayed or incorrect diagnoses can lead to prolonged recovery times, worsened conditions, and higher risks of complications. In recent years, machine learning and deep learning have emerged as powerful tools for medical image analysis, offering the potential to assist healthcare professionals by automating parts of the diagnostic process. This thesis concerns the development of a multi stage classification pipeline for the automated diagnosis of hip conditions from x-ray images, utilizing state-of-the-art deep learning techniques. The dataset used in this study includes a combination of publicly available hip x-ray images and additional images provided by a physician, covering fractured, operated, and healthy hips, as well as hips with osteoarthritis. The classification pipeline consists of five stages, each addressing a specific diagnostic task. These stages include whether the image shows the left or right hip, whether the hip is normal or not, if an operation has been performed, the type of operation (arthroplasty or nailing), and the classification of fracture types and conditions (intertrochanteric fracture, subcapital fracture, osteoarthritis). The final classification pipeline incorporates a ResNet50 model for the initial classification of left or right hip, achieving an accuracy of 89%. For the rest of the stages, VGG16 models were selected. The highest accuracy was obtained for classifying normal versus abnormal hips (98% recall) and the type of operation (100%). However, lower accuracy was observed in more complex tasks, such as differentiating between fracture types, where the model achieved an accuracy of 73%. Additionally, the classification where the hip is operated or not, the model achieved a recall of 91%. Transfer learning played a crucial role in boosting the performance of the pipeline, allowing the models to generalize well despite the limited availability of training data. Despite the effectiveness of the proposed pipeline, several limitations were encountered. One of the main challenges was the limited availability of open-source medical imaging data, which hindered the training of more robust models. Additionally, hardware limitations restricted the ability to train larger models or explore more complex architectures. Future research can also utilize generative models to synthesize additional medical images, expanding the training dataset and improving model performance. The findings of this thesis highlight the potential of deep learning techniques in automating medical diagnosis, particularly for hip-related conditions. Although automated diagnostic systems are still in the early stages of development, and should be used complementary to human expertise, they offer numerous benefits. These include faster and more efficient diagnosis, reduced diagnostic errors, and the ability to assist doctors in identifying additional areas of concern in medical images. Ultimately, automated systems could become valuable tools in healthcare, and drastically improve patient outcomes.