Non-Obstructive Azoospermia Prediction via Deep Learning-Driven Testicular Ultrasound Image Analysis: A Clinical Validation Study

Objectives To develop and validate a multimodal diagnostic model for non-obstructive azoospermia (NOA) by integrating deep learning (DL) analysis of testicular ultrasound images with clinical hormonal levels and ultrasonographic parameters. Methods This retrospective study analyzed 549 ultrasound images from 275 azoospermic patients (95 obstructive azoospermia(OA), 180 NOA) and 113 age-matched healthy controls. Conventional ultrasonographic and hemodynamic features were extracted. Five DL models were developed on manual testicular ROIs using five-fold cross-validation to differentiate OA from NOA, with Gradient-weighted Class Activation Mapping (Grad-CAM) employed for visualization. The optimal model's probability output was integrated with significant clinical and ultrasound features into a nomogram, which was validated on an independent prospective cohort of 100 patients. Results Among DL models, NasNetMobile model achieved the highest performance with an area under the curve (AUC) of 0.95 in the internal test set. The integrated nomogram, which combined the NasNetMobile probability with serum FSH, testicular volume, and EDV, demonstrated superior discriminative ability, attaining an AUC of 0.97 (95% CI: 0.91–1.00), accuracy of 96%, sensitivity of 97%, and specificity of 95% in the internal cohort, and maintained robustness in the prospective external validation (AUC = 0.97) (P<0.05). Grad-CAM provided visual confirmation of its focus on discriminative parenchymal regions. Conclusions The integration of DL-driven image analysis with clinical and ultrasonographic parameters into a nomogram provides a highly accurate, non-invasive tool for preoperative NOA diagnosis, potentially reducing the need for diagnostic biopsies.