[Paper Review] RRA-U-Net: a Residual Encoder to Attention Decoder by Residual Connections Framework for Spine Segmentation under Noisy Labels

Ziyang Wang, Zhengdong Zhang+1 more|arXiv (Cornell University)|Sep 27, 2020

Medical Imaging and Analysis19 references4 citations

TL;DR

This paper proposes RRA-U-Net, a U-Net-based framework with residual encoders, residual-block-based skip connections, attention modules, and an adaptive clean noisy label learning strategy (ACNLL) for spine segmentation under noisy labels. It achieves state-of-the-art performance on a public spine CT dataset by effectively mitigating feature distribution shifts and improving robustness to label noise.

ABSTRACT

Segmentation algorithms of medical image volumes are widely studied for many clinical and research purposes. We propose a novel and efficient framework for medical image segmentation. The framework functions under a deep learning paradigm, incorporating four novel contributions. Firstly, a residual interconnection is explored in different scale encoders. Secondly, four copy and crop connections are replaced to residual-block-based concatenation to alleviate the disparity between encoders and decoders, respectively. Thirdly, convolutional attention modules for feature refinement are studied on all scale decoders. Finally, an adaptive clean noisy label learning strategy(ACNLL) based on the training process from underfitting to overfitting is studied. Experimental results are illustrated on a publicly available benchmark database of spine CTs. Our segmentation framework achieves competitive performance with other state-of-the-art methods over a variety of different evaluation measures.

Motivation & Objective

• To address the challenge of training robust medical image segmentation models under noisy labels.
• To reduce feature distribution discrepancies between encoder and decoder features in U-Net architectures.
• To enhance feature refinement in decoders through convolutional attention modules.
• To develop a training strategy that adapts to label noise by leveraging the transition from underfitting to overfitting.
• To achieve state-of-the-art performance on spine CT segmentation despite label noise.

Proposed method

• Introduces residual interconnections across multiple scale encoders to stabilize feature learning and gradient flow.
• Replaces traditional copy-and-crop skip connections with residual-block-based concatenation to reduce feature distribution mismatch between encoder and decoder.
• Employs convolutional attention modules at all decoder stages to refine features and emphasize relevant anatomical structures.
• Proposes an adaptive clean noisy label learning (ACNLL) strategy that dynamically adjusts learning based on training dynamics from underfitting to overfitting.
• Trains the model end-to-end using a loss function that integrates attention-guided supervision and noise-aware optimization.
• Leverages a publicly available spine CT dataset to validate the framework under realistic noisy label conditions.

Experimental results

Research questions

• RQ1 How can residual connections across encoder scales improve feature representation in spine segmentation under label noise?
• RQ2 To what extent do residual-block-based skip connections reduce the feature distribution gap between encoder and decoder?
• RQ3 Can attention modules in all decoder stages enhance segmentation accuracy and robustness to noisy labels?
• RQ4 How effective is the ACNLL strategy in identifying clean samples during training under noisy label regimes?
• RQ5 Does the proposed framework outperform existing state-of-the-art methods on spine CT segmentation with noisy labels?

Key findings

• The RRA-U-Net framework achieves competitive segmentation performance across multiple evaluation metrics on a public spine CT benchmark.
• The use of residual-block-based skip connections significantly reduces the feature distribution discrepancy between encoder and decoder features.
• Attention modules in all decoder stages improve feature refinement and lead to more accurate boundary predictions.
• The ACNLL strategy effectively identifies and prioritizes clean samples during training, enhancing model generalization under noisy labels.
• The framework demonstrates robustness to label noise, maintaining high performance even when labels are imperfect.