[Paper Review] Divided We Stand: A Novel Residual Group Attention Mechanism for Medical Image Segmentation

Chaitanya Kaul, Nick Pears+1 more|arXiv (Cornell University)|Dec 4, 2019

AI in cancer detection22 references5 citations

TL;DR

This paper proposes FocusNetAlpha, a novel residual block integrating residual learning, group convolutions, and a new attention mechanism for medical image segmentation. By combining these components with a hybrid loss function, the model achieves state-of-the-art performance on ISIC 2018, cell nuclei, and DRIVE datasets, improving IoU by 5.6% over basic residual blocks and 1.3% over ResNeXt group convolution blocks, while using fewer parameters and FLOPs.

ABSTRACT

Given that convolutional neural networks extract features via learning convolution kernels, it makes sense to design better kernels which can in turn lead to better feature extraction. In this paper, we propose a new residual block for convolutional neural networks in the context of medical image segmentation. We combine attention mechanisms with group convolutions to create our group attention mechanism, which forms the fundamental building block of FocusNetAlpha - our convolutional autoencoder. We adapt a hybrid loss based on balanced cross entropy, tversky loss and the adaptive logarithmic loss to create a loss function that converges faster and more accurately to the minimum solution. On comparison with the different residual block variants, we observed a 5.6% increase in the IoU on the ISIC 2017 dataset over the basic residual block and a 1.3% increase over the resneXt group convolution block. Our results show that FocusNetAlpha achieves state-of-the-art results across all metrics for the ISIC 2018 melanoma segmentation, cell nuclei segmentation and the DRIVE retinal blood vessel segmentation datasets with fewer parameters and FLOPs. Our code and pre-trained models will be publicly available on GitHub to maximize reproducibility.

Motivation & Objective

• To improve feature extraction in medical image segmentation by designing a more effective convolutional kernel.
• To address the limitations of standard residual blocks and group convolution architectures in capturing complex anatomical structures.
• To enhance segmentation accuracy through a novel attention mechanism integrated with group convolutions.
• To accelerate convergence and improve optimization stability using a hybrid loss function combining balanced cross-entropy, Tversky loss, and adaptive logarithmic loss.
• To achieve state-of-the-art performance with reduced model complexity in terms of parameters and FLOPs.

Proposed method

• Proposes a new residual block that integrates residual learning with group convolutions and a novel attention mechanism.
• Introduces a group attention mechanism that learns to focus on relevant feature groups across channels and spatial locations.
• Employs a hybrid loss function combining balanced cross-entropy, Tversky loss, and adaptive logarithmic loss to improve training dynamics.
• Designs FocusNetAlpha as a U-Net-like autoencoder architecture built from the proposed residual blocks.
• Utilizes channel-wise and spatial-wise attention to dynamically reweight feature maps based on their relevance to segmentation tasks.
• Optimizes the loss function to be more robust to class imbalance and to converge faster to optimal solutions.

Experimental results

Research questions

• RQ1 Can a novel residual block combining group convolutions and attention mechanisms improve segmentation accuracy in medical imaging?
• RQ2 How does the proposed group attention mechanism compare to standard residual and ResNeXt blocks in terms of feature representation and segmentation performance?
• RQ3 To what extent does the hybrid loss function accelerate convergence and improve optimization stability in medical image segmentation?
• RQ4 Can the proposed architecture achieve state-of-the-art performance with fewer parameters and FLOPs compared to existing models?
• RQ5 Does the integration of attention within grouped convolutions lead to better generalization across diverse medical imaging tasks?

Key findings

• FocusNetAlpha achieves a 5.6% improvement in Intersection over Union (IoU) on the ISIC 2017 dataset compared to the basic residual block.
• The model shows a 1.3% IoU gain over the ResNeXt group convolution block on the ISIC 2017 dataset.
• FocusNetAlpha achieves state-of-the-art performance across all metrics on the ISIC 2018 melanoma segmentation, cell nuclei segmentation, and DRIVE retinal blood vessel segmentation datasets.
• The model achieves superior performance with fewer parameters and lower FLOPs compared to baseline architectures.
• The hybrid loss function enables faster and more accurate convergence to the optimal solution during training.
• The proposed method demonstrates strong generalization across diverse medical imaging tasks, including skin lesion, cell nuclei, and retinal vessel segmentation.