[Paper Review] MixModule: Mixed CNN Kernel Module for Medical Image Segmentation

Henry Yu, Feng Xue+2 more|arXiv (Cornell University)|Oct 19, 2019

Advanced Neural Network Applications26 references4 citations

TL;DR

This paper proposes MixModule, a novel convolutional neural network module that fuses multiple kernel sizes (3×3, 5×5, 7×7) to capture multi-scale contextual features in medical image segmentation. By replacing single-size kernels in U-Net and its variants (U-Net, R2U-Net, Attention U-Net), MixModule improves feature representation, achieving state-of-the-art performance across three medical image segmentation benchmarks with consistent gains in Dice score and other metrics.

ABSTRACT

Convolutional neural networks (CNNs) have been successfully applied to medical image classification, segmentation, and related tasks. Among the many CNNs architectures, U-Net and its improved versions based are widely used and achieve state-of-the-art performance these years. These improved architectures focus on structural improvements and the size of the convolution kernel is generally fixed. In this paper, we propose a module that combines the benefits of multiple kernel sizes and we apply the proposed module to U-Net and its variants. We test our module on three segmentation benchmark datasets and experimental results show significant improvement.

Motivation & Objective

• To address the limitation of fixed-size convolution kernels in medical image segmentation networks, which may fail to capture diverse receptive fields.
• To improve feature representation in U-Net and its variants by fusing multi-scale contextual information from different kernel sizes.
• To develop a plug-and-play module that enhances segmentation performance without architectural overhaul.
• To validate the effectiveness of the proposed module across diverse medical imaging tasks with limited data.

Proposed method

• MixModule employs parallel 3×3, 5×5, and 7×7 convolutional kernels applied to the same input feature map to capture multi-scale spatial features.
• The outputs of the multiple kernels are concatenated along the channel dimension to form a unified multi-receptive field representation.
• The module is designed as a plug-in replacement for standard convolutional blocks in U-Net and its variants, including R2U-Net and Attention U-Net.
• The method uses standard ReLU activation and batch normalization, with feature maps processed through concatenation to preserve spatial resolution.
• The architecture maintains the U-Net skip-connection design, ensuring high-resolution supervision during training.
• The module is trained end-to-end using Adam optimizer with learning rate decay and data augmentation (rotation, flip, brightness, contrast, shift).

Experimental results

Research questions

• RQ1 Can combining multiple kernel sizes improve feature representation in medical image segmentation networks?
• RQ2 Does the proposed MixModule consistently improve performance across different U-Net variants and datasets?
• RQ3 How does MixModule compare to standard single-kernel convolution blocks in terms of segmentation accuracy and robustness?
• RQ4 Can MixModule effectively handle small medical image datasets with limited annotations?

Key findings

• On the Skin lesion dataset, MixU-Net achieved a Dice score of 0.7673, outperforming standard U-Net (0.7645) and all other variants.
• In the DRIVE retinal vessel segmentation dataset, MixAttU-Net achieved a Dice score of 0.6820, surpassing standard AttU-Net (0.6617) and other baselines.
• On the CHASE_DB1 dataset, MixAttU-Net achieved a Dice score of 0.6636, exceeding standard AttU-Net (0.6457) and R2U-Net (0.6378).
• MixModule-based models consistently achieved the highest F1-score and Dice coefficient across all three datasets, indicating improved segmentation robustness.
• The MixU-Net model achieved 95.12% accuracy on the Skin dataset, compared to 94.96% for standard AttU-Net.
• The ablation results confirm that multi-scale kernel fusion significantly enhances feature learning, especially in capturing fine-grained structures like blood vessels and lesion boundaries.