Detection of hamstring injuries on football players with thermal imaging and artificial intelligence

Among the key factors influencing the success of football teams are player quality, tactical arrangements, and training programs. However, one of the primary factors negatively impacting this success is player injuries. An active injury prevents a player from contributing to the team, thereby reducing team performance and incurring additional costs for treatment. Consequently, football teams utilize various medical imaging methods, including ultrasonography, magnetic resonance imaging, computed tomography, blood tests, and isokinetic devices, to examine muscle injuries, enable early diagnosis, monitor rehabilitation processes, and determine return-to-play decisions. Recent advancements in thermography have introduced thermographic imaging as a cost-effective, portable, non-invasive, and user-friendly method for injury analysis in football teams. By utilizing temperature asymmetry and increased heat in the injured muscle region, thermography enables detailed examination of injuries. This has led to a rise in studies focusing on the use of thermography in sports injuries in recent years. This thesis investigates the use of thermographic imaging in analysing football player injuries and its applicability in injury detection and rehabilitation planning. The study utilizes deep learning methods to segment muscle regions from lower extremity thermographic images and classify injuries in these regions. Additionally, decision support systems for the detection of active injuries and planning of rehabilitation processes in football players were developed. For the research, lower extremity thermographic images were collected from football players in a Turkish Super League team over two seasons. Although large datasets are generally required to achieve successful results in deep learning methods, the limited number of injured players per season presents challenges in achieving balanced datasets. To overcome this issue, in addition to traditional deep learning methods, few-shot learning techniques and novel deep learning methods were employed. In this thesis, the thermographic images of football players with active injuries were analyzed, and their injuries were detected, followed by planning their rehabilitation processes. Subsequently, deep learning algorithms were utilized to segment muscle regions for more detailed analysis of these injuries, and the segmented regions were classified to identify injuries. To address the issue of insufficient and imbalanced datasets, few-shot learning and modern deep learning methods were employed to enhance classification accuracy. The implementations conducted within the scope of the thesis were carried out under three main stages. First, injuries in football players were detected from thermographic images, and the thermal response of injuries under exercise was analyzed. Additionally, thermography was used to assist in planning rehabilitation processes and determining return-to-play decisions. Subsequently, muscle regions identified based on the anatomic atlas were segmented using deep learning methods such as U-Net, Pyramid Scene Parsing Network, LinkNet, and Feature Pyramid Network. The segmentation results demonstrated that U-Net achieved the highest performance with a success rate of 99%. Following the segmentation, injuries in the segmented muscle regions were detected using deep learning methods such as DenseNet, Visual Geometry Group, ResNet, and EfficientNet. An end-to-end algorithm structure encompassing both segmentation and classification was designed for injury detection. The classification results indicated that the highest accuracy rates were achieved with EfficientNetB0 (83.9%) and EfficientNetB1 + Feature Pyramid Network (81.0%). To further enhance accuracy and address the issue of imbalanced datasets, classification was also performed using Siamese Networks, Prototypical Networks, and Kolmogorov-Arnold Network. The classification results showed that the Siamese and Prototypical Networks achieved success rates of 94% and 97.78%, respectively, while the Kolmogorov-Arnold Network achieved a success rate of 93.1%. This demonstrates the effectiveness of few-shot learning methods and the Kolmogorov-Arnold Network, a novel deep learning method, in detecting sports injuries using thermography. Furthermore, this study is among the first in the literature to use Kolmogorov-Arnold Networks for injury detection from thermographic images. In conclusion, the studies conducted within the scope of this thesis demonstrate the significant contribution of thermography and deep learning methods to sports medicine by enabling effective injury detection and management.