Interpretable Sperm Morphology Classification via A... | AI Research

Interpretable Sperm Morphology Classification via Attention-Guided Deep Learning
This research addresses the challenge of automating sperm morphology analysis, a critical but time-consuming process in diagnosing male infertility. While deep learning models can analyze medical images, they often function as "black boxes," providing results without explanation, which hinders their use in clinical settings. This study introduces an attention-guided framework that not only improves the accuracy of sperm classification but also provides visual evidence to help clinicians understand how the model reaches its decisions.

How the Approach Works

The researchers developed a framework that combines the EfficientNet-B0 model—a compact and efficient image processor—with a Convolutional Block Attention Module (CBAM). This attention module acts like a filter, forcing the model to prioritize the most diagnostically relevant parts of the sperm head while ignoring background noise.
To ensure the model performs well even on small datasets, the team implemented a "freeze-then-unfreeze" training strategy. In the first phase, they kept the core feature-extraction layers frozen to prevent overfitting, training only the classification head. In the second phase, they fine-tuned the entire model at a lower learning rate. Finally, they used a technique called Grad-CAM++ to generate heatmaps, which visually highlight the specific regions of the sperm that influenced the model's classification.

Key Findings

The proposed model was tested on two public datasets, SMIDS and HuSHem, and consistently outperformed both a custom SimpleCNN and a standard EfficientNet-B0. On the SMIDS dataset, the model achieved 90.2% accuracy. The improvement was even more pronounced on the smaller HuSHem dataset, where the model reached 93.9% accuracy, significantly higher than the baseline models.
The visual heatmaps generated by Grad-CAM++ confirmed that the model is "looking" at the right places. For example, when identifying abnormal sperm, the model focused on irregular boundaries and deformed regions, while for normal sperm, it focused on the smooth, oval structure of the head. This alignment with clinical criteria is essential for building trust in automated diagnostic tools.

Considerations for Clinical Use

The study highlights that while deep learning is powerful, its success depends on how it is adapted for specific medical tasks. The researchers found that simply applying a standard pretrained model to small medical datasets can lead to poor performance due to overfitting. By integrating attention mechanisms and a structured training strategy, they were able to overcome these limitations.
While the results are promising, the authors note that the HuSHem test set was relatively small, consisting of only 33 samples. They suggest that future research should focus on validating this framework using larger, multi-centric datasets to ensure the model remains robust and reliable across different laboratory environments.