Loss Function Selection for Vision Transformers in the Automatic Diagnosis of Bone Damage Based on X-ray Images with Low Variability in Training Data
摘要
Fracture diagnostics based on X-ray images is one of the most sought-after tasks in the field of automated medical research using artificial intelligence technologies. The aim of this article is an experimental study of the effectiveness of using several popular classes of vision transformers (ViT, variations of Swin and DeiT) in combination with five loss function variants (center focus, background suppression, single peak, attention dispersion, feature map) oriented to taking into account various features of medical images for solving the problem of fracture detection using elements of the transfer learning approach. A comparison was also made with the results of the ResNet-50 convolutional network. The quality of the results was assessed using standard classification metrics—F1-score and areas under the ROC curve. Two open datasets (almost 9800 images in total) with and without bone damage examples were used for testing. The first, smaller in size, contains information only about the clavicles, while the second, much larger in size, covers injuries to the limbs, lower back, hip, and knees. This made it possible to test elements of the transfer learning approach in the context of solving a medical diagnostic problem. Replacing the base loss function improved the quality of automatic damage detection—the metric of area under the ROC curve increased to 0.97 compared to the standard value of 0.73. In transfer learning, the best fine-tuned ViT model with feature map loss demonstrated values of the F1-score at the level of 0.85 with a base value of 0.9. The results achieved represent a foundation for the effective application of artificial intelligence technologies in medical diagnostics. Further research in this area may involve expanding the analyzed datasets and complicating their structure, including situations with even more significant class imbalances, as well as considering other types of neural network architectures and implementing trainable parameters in loss functions built on the basis of the approaches discussed in this article.