Filtern
Erscheinungsjahr
- 2021 (1)
Dokumenttyp
- Masterarbeit (1)
Sprache
- Englisch (1)
Volltext vorhanden
- ja (1)
Gehört zur Bibliographie
- ja (1)
Schlagworte
- Bilderkennung (1)
- Maschinelles Lernen (1)
Institut
Every fifth human suffers from musculoskeletal disorders. Diagnosis, treatment and aftercare of those conditions provide a severe problem for the healthcare system around the world. Radiology units are highly occupied with patient demands and in consequence, radiologists are especially prone to suffer from fatigue. Overall 11% to 27% fractures are misdiagnosed. This is especially harmful during childhood since inappropriate treatments are futile for further bone development. Certain care in this perspective is epiphyseal fractures, which are directly linked to bone growth.In recent years more and more computational driven diagnosis methods find their way into clinical diagnosis due to hardware and software advancements. Particularly the evolvement of machine learning algorithms for image analysis in any kind of application is intriguing for clinical usage to support radiologists. Here, we compare state-of-art imaging analysis machine learning modells for their applicability to predict epiphyseal fractures. Using 21557 X-ray images, split into training and validation datasets, from the Radiology Department of the Medical University of Graz were used to train seven different commonly used models. We found that A custom ensemble model of ShapeMask and SpineNet yielded the most accurate prediction with 80,7%. Most tested models, including SpineNet, ResNet, MobileNetV2, a custom convolutional neural network, faster R-CNN, and Azure Custom AI provided an accuracy of 70% to 80%. Only MobileNetV2 turned out to be inapplicable for this use case, resulting in a meagre 40% accuracy. None of the tested models was able to outperform the accuracy of radiologists. All in all, we provide a comprehensive overview of the possible utilization of currently available imaging analysis machine learning models and their possible use for epiphyseal fracture diagnosis.