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

Area of Science:

• Medical imaging
• Artificial intelligence in medicine
• Diagnostic tools

Background:

• Acute scrotum is a common pediatric surgical emergency.
• Timely diagnosis of testicular torsion is crucial to prevent testicular loss.
• Ultrasound (US) is the primary imaging modality for acute scrotum evaluation.

Purpose of the Study:

• To develop and evaluate a deep learning model for the differential diagnosis of acute scrotum using single ultrasound images.
• To assess the model's ability to differentiate testicular torsion from other causes of acute scrotum.
• To interpret the model's decision-making process using Class Activation Mapping.

Main Methods:

• A binary classification model using EfficientNet architecture was trained on 1172 Doppler US images from patients with acute scrotal pain.
• The dataset was split into 70% training and 30% validation sets.
• Data augmentation and class weighting were employed to handle class imbalance, and Class Activation Mapping was used for interpretability.

Main Results:

• The deep learning model achieved high diagnostic performance with 97% accuracy, 98% precision, 97% sensitivity, and 97% F1 score.
• Class activation mapping highlighted critical regions like absent testicular blood flow and whirlpool signs, aiding diagnosis.
• A prospective pilot study of 20 patients showed the system correctly identified torsion in confirmed cases.

Conclusions:

• Deep learning models show promise for diagnosing acute scrotal emergencies from single ultrasound images.
• Preliminary feasibility assessment in a small pilot study is encouraging.
• Larger, multicenter studies are necessary to establish clinical utility and workflow integration for this AI tool.