TLDR: MicroAUNet is a novel, lightweight deep learning model designed for accurate and real-time segmentation of colorectal polyps in colonoscopy images. It combines boundary-enhanced multi-scale feature fusion with a two-stage knowledge distillation process, allowing it to achieve state-of-the-art accuracy with significantly reduced computational complexity, making it ideal for clinical applications.
Colorectal cancer remains a significant global health concern, and early detection and removal of precancerous polyps during colonoscopy are crucial for reducing mortality. However, traditional colonoscopy relies heavily on physician expertise, often leading to missed polyps and operational burdens. Deep learning-based computer-aided systems offer a promising solution, with accurate polyp image segmentation being a foundational step for diagnosis.
Current deep learning models for polyp segmentation face two main challenges: providing ambiguous polyp margins, which can compromise clinical decision-making, and relying on heavy architectures with high computational complexity, making them too slow for real-time endoscopic applications. Addressing these limitations, researchers have introduced MicroAUNet, a novel, lightweight, attention-based segmentation network.
MicroAUNet’s Innovative Design
MicroAUNet is designed to achieve both high boundary precision and computational efficiency. It incorporates several key innovations:
- Boundary-Enhanced Multi-scale Feature Fusion: This module combines depthwise separable dilated convolutions with a single-path, parameter-shared channel–spatial attention block. Depthwise separable dilated convolutions efficiently expand the network’s receptive field without adding many parameters, helping to capture multi-scale contextual information. The shared attention mechanism enhances feature discrimination by highlighting important channels and spatial regions, crucial for accurate boundary localization in complex endoscopic images, all while minimizing computational overhead.
- Progressive Two-Stage Knowledge Distillation: Lightweight networks often struggle with detailed semantic and boundary modeling due to their limited parameters. To overcome this, MicroAUNet employs a two-stage knowledge distillation framework. In this process, a high-capacity ‘teacher’ model (MALUNet) transfers its rich semantic and boundary knowledge to the smaller ‘student’ model (MicroAUNet).
The first stage, called Imitation Learning, focuses on aligning the student’s features and output predictions with the teacher’s. The second stage, Preference Alignment, uses the teacher’s high-confidence predictions as positive examples and low-confidence predictions as negative ones. This helps the student model refine its decision boundaries, especially in challenging scenarios, by encouraging consistency within classes and separation between them.
Performance and Efficiency
Extensive experiments were conducted on two public polyp segmentation datasets, Kvasir-SEG and CVC-ClinicDB. MicroAUNet was compared against several state-of-the-art models, including UNet, SANet, UNeXt, and MALUNet. The results demonstrated that MicroAUNet achieves an excellent balance between segmentation accuracy and efficiency, often outperforming other models across both datasets.
Notably, MicroAUNet achieved high mDice and mIoU scores (common metrics for segmentation accuracy) while using significantly fewer parameters. For instance, MicroAUNet has only 0.0249 million parameters, a substantial reduction compared to UNet (7.77 million) and SANet (23.90 million). This low model complexity makes it highly suitable for real-time clinical deployment.
Ablation studies, where individual components of MicroAUNet were removed, confirmed the importance of each design choice. Removing the depthwise separable dilated convolutions, the imitation learning stage, or the preference alignment stage all led to a noticeable decrease in performance, highlighting the synergistic effects of these components.
Also Read:
- Advancing Medical Image Analysis with Adaptable Foundation Models
- DINO-MX: A Flexible Framework for Self-Supervised Learning in Medical Imaging
Conclusion and Future Directions
MicroAUNet represents a significant step forward in colonoscopy polyp image segmentation. By integrating boundary-enhanced multi-scale fusion with a progressive two-stage knowledge distillation strategy, it delivers high accuracy with substantially reduced computational costs, making it a strong candidate for real-time clinical applications. While the model shows robust performance, future research will explore its generalization to unseen clinical environments and investigate self-distillation or teacher-free strategies to improve adaptability. The code for MicroAUNet is publicly available for further exploration and development. You can find more details in the full research paper: MicroAUNet: Boundary-Enhanced Multi-scale Fusion with Knowledge Distillation for Colonoscopy Polyp Image Segmentation.
