Deep learning-based classification of acute scrotum using single ultrasound images.

OBJECTIVE: To develop a deep learning model for differential diagnosis of acute scrotum using single ultrasound (US) images. PATIENTS AND METHODS: We analysed 1172 patients with acute scrotal pain evaluated by Doppler US at four hospitals. From each case, we selected a representative axial colour Doppler US image. We trained a binary classification model to distinguish torsion from non-torsion using an EfficientNet architecture. The dataset was split 70% for training and 30% for validation. We addressed class imbalance with data augmentation and class weighting. Class Activation Mapping was used to interpret model decisions. RESULTS: The model achieved robust performance: accuracy 97%, precision 98%, sensitivity 97%, and F1 score 97%. Class activation mapping heatmaps localised decision-making to pathologically critical regions, including absent testicular blood flow and whirlpool signs. In a 20-patient prospective pilot study, the system correctly identified both surgically confirmed torsion cases, with one non-torsion case misclassified as torsion. CONCLUSIONS: A deep learning model demonstrated promising diagnostic performance in differentiating acute scrotal emergencies using single US images. Its feasibility was preliminarily assessed in a small pilot study. These findings support further investigation, with larger and more balanced multicentre studies required to establish clinical utility and effective workflow integration.