Interpretable Multi-Label Classification for Tibiofibula Fracture 2D CT Images with Selective Attention and Data Augmentation

Abstract

Background: Tibiofibula fractures occur across all age groups, and postoperative complications are frequent. An accurate and rapid classification methodology for these fractures could significantly assist physicians. Clinically, tibiofibula fractures occur at various locations, and the fracture types are not evenly distributed. Methods: This paper presents a deep learning model for the interpretable multi-label classification of tibiofibula fractures in two-dimensional (2D) CT scan images, addressing the challenges posed by a limited sample size and an uneven distribution of fracture types. We retrospectively collected 2494 2D CT images from 168 patients with tibia or fibula fractures. The types of fractures identified in the CT scan images were classified according to the AO/OTA fracture classification. A deep learning model was developed to classify composite fractures in 2D CT images, providing visual interpretation for each identified class. The visual interpretation was given with the saliency maps constructed by the Grad-CAM++ method. The deep learning model was trained using data augmentation techniques to address class imbalance and the limited dataset size. Results: Our experiments demonstrated that the proposed model achieved a mean average precision (mAP) of 95.71%. Conclusions: The saliency map-based visual interpretation enables the verification of whether the model provides reliable decision-making for classification.