Detection and Segmentation of Heterogeneous Bone Tumours in Limited Radiographs

Bone tumours are a rare and often highly malignant entity. Early clinical diagnosis is the most important step, but the difficulty of detecting and assessing bone malignancies is in its radiological peculiarity and limited experience of non-experts. Since X-ray imaging is the first imaging method of bone tumour diagnostics, the purpose of this study is to develop an artificial intelligence (AI) model to detect and segment the tumorous tissue in a radiograph. We investigated which methods are necessary to cope with limited and heterogeneous data. We collected 531 anonymised radiographs from our musculoskeletal tumour centre. In order to adapt to the complexity of recognizing the malignant tissue and cope with limited data, transfer learning, data augmentation as well as several architectures, some of which were initially designed for medical images, were implemented. Furthermore, dataset size was varied by adding another bone tumour entity. We applied a data split of 72%, 18%, 10% for training, validation and testing, respectively. To provide statistical significance and robustness, we applied a cross-validation and image stratification with respect to tumour pixels present. We achieved an accuracy of 99.72% and an intersection over union of 87.43% for hold-out test data by applying several methods to tackle limited data. Transfer learning and additional data brought the greatest performance increase. In conclusion, our model was able to detect and segment tumorous tissue in radiographs with good performance, although it was trained on a very limited amount of data. Transfer Learning and data augmentation proved to significantly mitigate the issue of limited data samples. However, to accomplish clinical significance, more data has to be acquired in the future. Through minor adjustments, the model could be adapted to other musculoskeletal tumour entities and become a general support tool for orthopaedic surgeons and radiologists.