TLDR: A new method called interpretability-aware pruning reduces the size of deep learning models used in medical image analysis while maintaining accuracy and making them more transparent. It achieves this by using techniques like Layer-wise Relevance Propagation (LRP) to identify and remove less important parts of the neural network, leading to efficient and understandable AI systems for healthcare.
Deep learning models have brought about significant advancements in medical image analysis, helping with tasks like disease detection and diagnosis. However, their widespread use in hospitals and clinics faces two major hurdles: their large size, which requires substantial computing power, and their ‘black-box’ nature, meaning it’s hard to understand how they make decisions. This lack of transparency can hinder trust among medical professionals who need clear insights into AI-driven diagnoses.
To address these challenges, researchers have developed a new method called interpretability-aware pruning. This innovative framework aims to reduce the complexity of these powerful AI models while ensuring they maintain their high predictive performance and, crucially, their transparency. By selectively keeping only the most important parts of each layer within the neural network, this method allows for targeted compression that preserves clinically meaningful representations.
Traditional pruning techniques often remove parts of a model based on simple rules, like how small a weight is. While effective for reducing size, these methods might accidentally discard components vital for the model’s decision-making or its ability to be interpreted. The interpretability-aware pruning approach, however, uses advanced ‘attribution techniques’ to guide the compression process. These techniques, including Layer-wise Relevance Propagation (LRP), DL-Backtrace (DLB), and Integrated Gradients (IG), calculate ‘importance scores’ for individual layers and neurons, showing how much each contributes to the model’s prediction. These scores then inform a precise pruning strategy, allowing the removal of less relevant components while maintaining performance and enhancing transparency.
The methodology involves several key steps. First, importance scores are computed for individual neurons using the chosen attribution methods. To make sure these scores are reliable, a small subset of samples (10 per class) from the training data is used. Different sampling techniques were explored, such as selecting samples the model is most confident about, random selection, or clustering-based sampling to ensure diversity. Once scores are obtained, neurons with the lowest importance are removed up to a specified pruning threshold. This effectively eliminates their contribution, leading to a more compact model.
Experiments were conducted across multiple medical image classification benchmarks using four diverse datasets: MURA (musculoskeletal X-rays), KVASIR (endoscopic images), CPN (common peroneal nerve injuries), and FETAL Planes (fetal ultrasound images). The pruning methodology was applied to three widely used deep learning architectures: VGG19, ResNet50, and Vision Transformer (ViT). The results consistently showed that this interpretability-guided pruning allows for high model compression with minimal loss in accuracy, often within a 5% margin, even at significant pruning rates.
Also Read:
- Advancing Cancer Prognosis: A New Framework for Generalizing Multimodal Models Across Cancer Types
- A New Approach to Combat Catastrophic Forgetting in AI: Focusing on the Final Learning Stages
Key Findings from the Research
Among the tested architectures, Vision Transformers (ViTs) demonstrated exceptional resilience, often preserving performance even after 65–80% of their neurons were pruned, and in some cases, up to 85%. This suggests that ViTs might have more redundancy or richer features compared to Convolutional Neural Networks (CNNs). VGG architectures typically tolerated up to 45–55% pruning with less than a 5% accuracy drop, while ResNets were more sensitive, maintaining stability only up to 25–35% pruning.
Regarding the interpretability methods, Layer-wise Relevance Propagation (LRP) consistently achieved the highest pruning thresholds while maintaining accuracy, slightly outperforming Integrated Gradients (IG) and DL-Backtrace (DLB). This highlights LRP’s effectiveness in identifying less important neurons. For sampling strategies, clustering-based sampling generally performed best, especially at moderate pruning levels, followed by confidence-based sampling. Interestingly, in some scenarios, a slight increase in accuracy was observed after pruning. This counterintuitive result suggests that removing noisy or negatively influential neurons can sometimes lead to better generalization by the model.
The implications for healthcare are significant. This interpretability-aware pruning framework enables the creation of lightweight models with reduced memory and computational needs. This means AI diagnostics can be deployed on edge devices, mobile platforms, and in clinics with limited resources, speeding up diagnostic workflows. Furthermore, by linking model compression with transparency, the method helps clinicians understand which neurons are critical, fostering trust in AI decisions based on meaningful features. The use of diverse sampling methods also ensures that the pruning process preserves neurons important for varied inputs, enhancing the model’s overall reliability.
In conclusion, this research introduces a principled and transparent approach to model compression for medical image analysis. By leveraging interpretability techniques, it paves the way for efficient, interpretable, and potentially more robust AI systems, which is crucial for the safe and trustworthy deployment of AI in high-stakes healthcare environments. For more details, you can refer to the full research paper here.
