Sperm Morphology Classification Using Xception-CBAM: A Deep Learning Approach on the SMIDS Dataset

Abstract

Accurate classification of sperm morphology is fundamental in evaluating male fertility and supporting reproductive health diagnostics. However, conventional manual assessment methods are often subjective, labor-intensive, and prone to inconsistencies. To address these limitations, this study presents a deep learning-based framework that integrates the Xception convolutional neural network with a Convolutional Block Attention Module (CBAM) to enhance automated classification performance. The model is trained and evaluated on the Sperm Morphology Image Dataset (SMIDS), comprising 3,000 high-resolution microscopic images categorized into Abnormal Sperm, Normal Sperm, and Non-Sperm classes. By leveraging transfer learning from ImageNet and incorporating both spatial and channel attention mechanisms, the model selectively emphasizes diagnostically salient features while suppressing irrelevant information. Experimental results demonstrate high generalization capability, achieving a test accuracy of 96.2%, with macro-averaged precision, recall, and F1-score of 95.0%, 95.3%, and 95.1%, respectively. The average Area Under the ROC Curve (AUC) reached 0.99 across all classes. Additional analyses, including confusion matrix evaluation, ROC and precision–recall curves, and class-wise performance metrics, confirm the model’s robustness and clinical reliability. Qualitative assessments further validate its discriminative power in real-world scenarios. This research underscores the potential of attention-augmented convolutional architectures in medical image analysis and offers a scalable, interpretable, and efficient tool for sperm morphology assessment in clinical and laboratory environments.

Keywords: Sperm Morphology Classification, Medical Image Analysis, Xception Network, CBAM Attention Mechanism, Deep Learning, SMIDS Dataset, Male Infertility Diagnosis, Computer-Aided Diagnosis.