Interpretable Sperm Morphology Classification via Attention-Guided Deep Learning

Technical Summary: Interpretable Sperm Morphology Classification via Attention-Guided Deep Learning

Problem Statement
Male infertility, affecting nearly 15% of couples globally, is frequently linked to abnormal sperm morphology. While semen analysis is the primary diagnostic tool, current manual assessment by laboratory specialists is time-consuming, subjective, and prone to low inter-observer agreement (60–70%). Although deep learning (DL) offers automated alternatives, most existing models function as "black boxes," lacking the interpretability required for clinical adoption. The core challenge addressed is the need for a system that not only achieves high classification accuracy but also provides transparent, clinically verifiable decision-making processes, particularly when working with limited medical datasets.

Methodology
The authors propose an attention-guided deep learning framework designed to balance performance with explainability. The pipeline consists of four stages:

1. Architecture: The model utilizes a pretrained EfficientNet-B0 backbone for feature extraction, enhanced by a Convolutional Block Attention Module (CBAM). CBAM sequentially applies channel and spatial attention mechanisms to force the network to focus on diagnostically critical regions of the sperm head, suppressing background noise and artifacts.
2. Training Strategy: To address the scarcity of labeled medical data (specifically the small HuSHem dataset), a freeze-then-unfreeze strategy is employed. Initially, the backbone and CBAM layers are frozen while only the classification head is trained. Subsequently, all parameters are jointly fine-tuned with a lower learning rate.
3. Data Augmentation: Standard augmentations (flipping, rotation, color jittering) are used. For the small HuSHem dataset, MixUp augmentation and label smoothing are applied as regularization strategies to prevent overfitting.
4. Explainability: Grad-CAM++ is integrated to generate class-specific visual heatmaps from the final convolutional layers, highlighting the specific image regions influencing the model's predictions.

Key Contributions
The paper outlines three primary contributions:

• Framework Proposal: An EfficientNet-B0 + CBAM architecture specifically designed for accurate and interpretable sperm morphology classification.
• Training Optimization: The integration of a freeze-then-unfreeze training strategy, demonstrated to be effective for enhancing performance on small-scale medical datasets.
• Clinical Transparency: The generation of visual explanations via Grad-CAM++ to validate that the model's decision-making aligns with clinical criteria, thereby fostering trust.

Experimental Results
The framework was evaluated on two public datasets: SMIDS (3,000 images, 3 classes) and HuSHem (216 images, 4 classes). The proposed model was compared against a baseline SimpleCNN and a standard pretrained EfficientNet-B0.

• SMIDS Performance: The proposed model achieved 90.2% accuracy and a macro F1-score of 0.913, outperforming both SimpleCNN (82.67%) and standard EfficientNet-B0 (88.00%).
• HuSHem Performance: The improvement was more pronounced on the smaller dataset. The proposed model reached 93.9% accuracy and an F1-score of 0.948. In contrast, the standard EfficientNet-B0 dropped to 63.64% accuracy, and SimpleCNN achieved 72.73%.
• Ablation Study: The study revealed that while a standard pretrained backbone alone performed worse than a simple baseline on the small HuSHem dataset (due to overfitting on background variations), the addition of the CBAM module improved accuracy by 16.16%. The freeze-then-unfreeze strategy contributed an additional 9.60% gain.
• Visual Validation: Grad-CAM++ visualizations confirmed that the model focused on clinically relevant sperm head structures (e.g., smooth oval shapes for normal sperm, irregular boundaries for abnormal ones) rather than background artifacts.

Significance and Claims
The authors claim that this work demonstrates that accurate and transparent frameworks are viable tools for automated sperm analysis in fertility clinics. The study highlights a critical lesson for medical AI: standard pretrained networks can underperform on limited data due to overfitting, but integrating spatial/channel attention mechanisms with targeted fine-tuning strategies can reverse this trend. By ensuring the model's focus aligns with manual clinical criteria, the proposed framework addresses the "black box" limitation, making transparency a core requirement rather than an afterthought. This approach is presented as a practical step toward deploying automated tools in real-world fertility laboratories.