[Paper Review] CaraNet: Context Axial Reverse Attention Network for Segmentation of Small Medical Objects

Ange Lou, Shuyue Guan+2 more|arXiv (Cornell University)|Aug 16, 2021

Radiomics and Machine Learning in Medical Imaging4 citations

TL;DR

CaraNet is a novel convolutional neural network that enhances small medical object segmentation through a Context Axial Reverse Attention mechanism, improving feature representation by focusing on contextual and long-range dependencies. It achieves state-of-the-art mean Dice scores on BraTS 2018 and multiple polyp datasets, demonstrating superior performance on small structures critical for early disease detection.

ABSTRACT

Segmenting medical images accurately and reliably is important for disease diagnosis and treatment. It is a challenging task because of the wide variety of objects' sizes, shapes, and scanning modalities. Recently, many convolutional neural networks (CNN) have been designed for segmentation tasks and achieved great success. Few studies, however, have fully considered the sizes of objects, and thus most demonstrate poor performance for small objects segmentation. This can have a significant impact on the early detection of diseases. This paper proposes a Context Axial Reserve Attention Network (CaraNet) to improve the segmentation performance on small objects compared with several recent state-of-the-art models. We test our CaraNet on brain tumor (BraTS 2018) and polyp (Kvasir-SEG, CVC-ColonDB, CVC-ClinicDB, CVC-300, and ETIS-LaribPolypDB) segmentation datasets. Our CaraNet achieves the top-rank mean Dice segmentation accuracy, and results show a distinct advantage of CaraNet in the segmentation of small medical objects.

Motivation & Objective

• To address the poor performance of existing CNNs in segmenting small medical objects due to limited context and feature refinement.
• To improve segmentation accuracy for small structures like brain tumors and polyps, which are critical for early diagnosis.
• To design a network architecture that effectively captures long-range spatial dependencies and enhances feature representation in low-resolution regions.
• To outperform state-of-the-art models on benchmark datasets for small object segmentation in medical imaging.
• To validate the effectiveness of axial and reverse attention mechanisms in enhancing feature learning for small object segmentation.

Proposed method

• Introduces a Context Axial Reverse Attention (CARA) module that applies axial attention along both height and width dimensions to capture long-range dependencies.
• Employs a reverse attention mechanism that suppresses irrelevant features and enhances discriminative features in low-resolution regions.
• Integrates the CARA module into an encoder-decoder U-Net style architecture to preserve spatial details and context.
• Utilizes multi-scale feature fusion to aggregate features from different stages, improving representation of small objects.
• Applies skip connections with learnable attention weights to refine feature maps at each decoder level.
• Trains end-to-end using binary cross-entropy and Dice loss to optimize segmentation performance.

Experimental results

Research questions

• RQ1 Can axial attention mechanisms improve feature representation for small medical objects in segmentation tasks?
• RQ2 Does a reverse attention mechanism enhance the model's ability to focus on small, low-contrast structures?
• RQ3 How does the proposed CaraNet compare to state-of-the-art models in terms of Dice score on small object segmentation benchmarks?
• RQ4 Can the integration of context-aware attention improve performance on diverse medical imaging modalities?
• RQ5 What is the contribution of axial and reverse attention to feature refinement in low-resolution regions?

Key findings

• CaraNet achieves the top-rank mean Dice score on the BraTS 2018 dataset, outperforming existing state-of-the-art models.
• On the Kvasir-SEG, CVC-ColonDB, CVC-ClinicDB, CVC-300, and ETIS-LaribPolypDB datasets, CaraNet achieves the highest mean Dice scores across all evaluated models.
• The model demonstrates a significant performance gain on small objects, with Dice scores increasing by up to 5.2% compared to baseline models on polyp datasets.
• Ablation studies confirm that both the axial and reverse attention components contribute independently to improved segmentation accuracy.
• The model maintains robustness across diverse imaging modalities and anatomical structures, including brain tumors and colon polyps.
• The results validate that context-aware attention mechanisms are essential for enhancing feature representation in small object segmentation.