Prediction of Sperm Retrieval Outcomes Based on Testicular Ultrasound Images and Dense Convolutional Sparse Coding

Assessment of sperm retrieval outcomes is important in assisted reproduction for non-obstructive azoospermia (NOA). Traditional assessment methods rely on microscopic observation or surgical sampling, which are invasive, highly operator-dependent and subjective. In recent years, the integration of medical imaging and artificial intelligence has offered new approaches for non-invasive diagnosis. However, research on automated analysis of testicular ultrasound images remains scarce. Currently, most studies are confined to clinical indicators or microscopic imaging, with a lack of systematic exploration and modeling of potential structural features within ultrasound images. To realize preoperative non-invasive assessment of sperm retrieval outcomes with ultrasound images, a multi-layer dense convolutional sparse coding (DCSC) network is proposed in this study. Firstly, the testicular region is segmented using 3D Slicer, and quantitative features are extracted from the segmented images via PyRadiomics. Subsequently, the features are input into the DCSC model, which employs an iterative soft thresholding algorithm (ISTA) to efficiently optimise sparse representations. This enables rapid extraction of key features while suppressing interfering information. Channel weighting is then performed using an enhanced squeeze-stimulate module. Finally, a dataset comprising 1,014 testicular ultrasound images is used to demonstrate the effectiveness of the proposed model. After multiple rounds of testing, the DCSC model achieved an area under the curve (AUC) value ranging from 0.8285 to 0.8520 on the validation set and from 0.8093 to 0.8254 on the test set, with the accuracy of 0.7783 to 0.7980 and 0.7586 to 0.7635, respectively. These results significantly outperform traditional methods such as convolutional neural networks (CNN) and random forests (RF).