Texture analysis of testicular ultrasound in infertile and fertile males

Purpose: Male infertility is a multifactorial pathology, and there are radiological methods that assist in addition to the gold standard diagnostic method, the spermiogram. Our aim is to measure the effectiveness of combining testicular ultrasound images obtained with the machine learning algorithms to predict spermiogram parameters. Materials and Methods: In this retrospective study, 169 cases who underwent testicular ultrasound and spermiogram examination within one month at our institution were included. Patients were divided into two groups: below 5th percentile and 5th-95th percentile, based on spermogram parameters including Sperm concentration, Total sperm count, Total motility, Progressive motility, Morphology, and TZI index. Two-dimensional segmentation of the testicular long segment and extraction of radiomics features were performed by a single physician using the "3D Slicer" program. Testicular parenchyma was manually segmented. Radiomics outputs consisted of original, fine-coarse Laplacian of Gaussian, and wavelet-transformed filtered images. A total of 1702 radiomics features were obtained for both testes. Voxel resampling was standardized to 1x1x1 mm³ for machine learning. Orange Data Mining software was used for Feature Selection. Feature selection methods applied were relieff, X2, Gain ratio, and Information gain. Machine learning algorithms used for Feature Classification were k-nearest neighborhood (kNN), decision tree, random forest, logistic regression, support vector machine (SVM), naive bayes, neural network, Adaboost, and Gradient boost. Parameters such as area under the curve (AUC), sensitivity, specificity, receiver operating characteristic (ROC), and confusion matrix were used to compare classifications. A p-value <0.05 was considered significant for all statistical results. Results: Naive Bayes model achieved the highest success in predicting the below 5th percentile group for Sperm concentration (AUC 0.77, sensitivity 0.59, specificity 0.75). Naive Bayes model also achieved the highest success in predicting the below 5th percentile group for Total sperm count (AUC 0.77, sensitivity 0.60, specificity 0.76). In predicting the below 5th percentile group for Total sperm motility, Naive Bayes model achieved the highest success (AUC 0.81, sensitivity 0.65, specificity 0.82). Naive Bayes model also obtained the highest success in predicting the below 5th percentile group for Progressive sperm motility (AUC 0.80, sensitivity 0.55, specificity 0.80). For predicting the below 5th percentile group for Sperm morphology, Naive Bayes model achieved the highest success (AUC 0.73, sensitivity 0.64, specificity 0.65). SVM model achieved the highest success in predicting the below 5th percentile group for Sperm TZI index (AUC 0.73, sensitivity 0.54, specificity 0.81). Additionally, comparisons were made between testis volume and reflux seconds measured after the Valsalva maneuver and the below 5th percentile groups. Strong negative correlations were found between these parameters. Statistical differences were also detected between sperm concentration and Total sperm count groups based on testicular echogenicity classifications (p<0.05). Conclusion: Our study showed that using parenchymal segmentation data obtained from testicular ultrasound, machine learning algorithms can provide insights into predicting below 5th percentile groups of spermiogram parameters. Keywords: Male infertility, testicular ultrasound, spermogram, machine learning, radiomics, high-level tissue features