Acute appendicitis CT imaging diagnosis system based on an improved nnU-Net

Acute appendicitis (AA) CT diagnosis has long faced the clinical challenge of low efficiency and high misdiagnosis rates in primary healthcare settings. To address this problem, we have developed an intelligent diagnostic system based on an improved nnU-Net, incorporating three key innovations: (1) a dynamically weighted composite loss function that combines Dice and cross-entropy losses and adjusts weights according to the training process, effectively improving segmentation accuracy for small appendiceal targets; (2) an edge-enhanced supervision mechanism that strengthens the model's perception of appendiceal boundaries through edge information; and (3) the use of Shapley Additive exPlanations (SHAP) to quantify the impact of key morphological features on diagnostic decisions, thereby enhancing system interpretability. We trained the system on CT data from 60 clinically confirmed acute appendicitis patients and evaluated it on an independent test set of 30 cases (15 with appendicitis, 15 normal). The system achieved a Dice coefficient of 72.4% in appendiceal segmentation. In terms of diagnostic performance, the AI system (sensitivity 73.3%, specificity 80.0%, accuracy 76.7%) performed comparably to senior physicians. Moreover, the system achieved an average diagnostic time of only 23.5 seconds, significantly improving efficiency. These findings suggest that our new AI system offers accuracy, speed, and interpretability, and has broad clinical application potential.