PaperlayerView exact revision on arXiv

Source-linked AI summary

Boundary-aware Context Neural Network for Medical Image Segmentation

Ruxin Wang, Shuyuan Chen, Chaojie Ji, Jianping Fan, Ye Li

arXiv:2005.00966v1eess.IVcs.CV

TL;DR

Medical image segmentation needs accurate boundaries despite lesion variation, low contrast, noise, and limited contextual or discriminative information in existing CNN-based methods. BA-Net uses boundary extraction, multi-task boundary and mask learning with interactive attention, and cross-level feature fusion in an encoder-decoder network. Evaluations on five public datasets report state-of-the-art or consistently improved segmentation performance across multiple tasks.

  • Problem

    Existing CNN-based medical segmentation methods can produce inaccurate object boundaries because context and discriminative feature information are limited, while medical images contain variation, indistinction, and noise.

  • Method

    BA-Net combines pyramid edge extraction, joint mask-and-boundary learning with interactive attention, and cross-level feature fusion in an encoder-decoder architecture.

  • Results

    BA-Net achieves state-of-the-art segmentation performance across skin lesion, colorectal polyp, lung, and optic disc tasks and shows consistent improvements on multiple public datasets.

  • Takeaways & Limitations

    Boundary cues and selectively aggregated multi-level features provide the basis for richer context and fine-grained features in the proposed 2D segmentation network.

Abstract

from arXiv · show

Medical image segmentation can provide a reliable basis for further clinical analysis and disease diagnosis. The performance of medical image segmentation has been significantly advanced with the convolutional neural networks (CNNs). However, most existing CNNs-based methods often produce unsatisfactory segmentation mask without accurate object boundaries. This is caused by the limited context information and inadequate discriminative feature maps after consecutive pooling and convolution operations. In that the medical image is characterized by the high intra-class variation, inter-class indistinction and noise, extracting powerful context and aggregating discriminative features for fine-grained segmentation are still challenging today. In this paper, we formulate a boundary-aware context neural network (BA-Net) for 2D medical image segmentation to capture richer context and preserve fine spatial information. BA-Net adopts encoder-decoder architecture. In each stage of encoder network, pyramid edge extraction module is proposed for obtaining edge information with multiple granularities firstly. Then we design a mini multi-task learning module for jointly learning to segment object masks and detect lesion boundaries. In particular, a new interactive attention is proposed to bridge two tasks for achieving information complementarity between different tasks, which effectively leverages the boundary information for offering a strong cue to better segmentation prediction. At last, a cross feature fusion module aims to selectively aggregate multi-level features from the whole encoder network. By cascaded three modules, richer context and fine-grain features of each stage are encoded. Extensive experiments on five datasets show that the proposed BA-Net outperforms state-of-the-art approaches.

I. INTRODUCTION

Medical image segmentation supports clinical analysis and diagnosis but remains difficult because lesions vary in appearance, contrast is often low, and CNNs can lose boundary detail and discriminative information. BA-Net addresses these challenges with boundary-aware, multi-task, and cross-level feature processing, achieving strong performance across several segmentation tasks.

  • Accurate lesion segmentation provides a basis for clinical analysis, disease diagnosis, and therapy planning.
  • Lesion segmentation is challenging because lesion size and shape vary, lesion-background contrast is low, and surrounding tissues can be complex.
  • Encoder-decoder CNNs learn semantic features through convolution and recover prediction masks through decoding, but successive operations can limit spatial detail.
  • BA-Net combines pyramid edge extraction, mini multi-task learning with interactive attention, and cross feature fusion within an encoder-decoder architecture.
  • The method selectively aggregates cross-level features and uses boundary information to provide complementary cues for segmentation prediction.
  • BA-Net achieves state-of-the-art segmentation performance across skin lesion, colorectal polyp, lung, and optic disc tasks.

II. RELATED WORK

Medical image segmentation research has progressed from handcrafted feature methods to deep CNN encoder-decoder architectures. Related approaches use intensity, texture, atlas, multi-scale, adversarial, and collaborative strategies to improve segmentation.

  • Conventional segmentation methods are commonly categorized as gray level-based, texture-based, and atlas-based approaches.
  • Gray level-based methods use strategies such as histogram statistics, edge detection, and region growing for segmentation.
  • Texture- and shape-prior methods have been applied to tasks including kidney segmentation in ultrasound images.
  • CNN-based encoder-decoder networks encode image content through convolutional layers and generate prediction masks through upsampling layers.
  • Recent methods enhance segmentation using dense deconvolution, chained residual pooling, auxiliary supervision, adversarial multi-scale losses, or collaborative architectures.

III. METHODOLOGY

BA-Net uses an encoder-decoder design with three cascaded modules to enrich context, extract multigranularity boundaries, and preserve fine spatial information. Its PEE, mini-MTL, and CFF components process features across encoder stages before sequential decoder prediction.

  • Overview: BA-Net uses a ResNet-based encoder-decoder with atrous convolutions, removed pooling in later blocks, and ASPP for multiscale feature encoding.The four residual-block outputs have spatial scales 1/4, 1/8, 1/8, and 1/8 of the input; ASPP combines parallel atrous convolutions and global average pooling.
  • Cross Feature Fusion Module: CFF selectively aggregates multi-level features from other encoder stages into the current stage, realizing information complementation across feature levels.The resulting features are combined with ASPP outputs in the decoder to produce sequential final segmentation predictions.
  • Pyramid Edge Extraction Module: The PEE module extracts multigranularity edge features by subtracting average-pooled responses at different sizes from local convolutional feature maps.It first applies 1×1 convolutions to stage features, then integrates the resulting edge maps with current-stage features through concatenation and a 1×1 convolution.
  • Pyramid Edge Extraction Module: PEE concatenates pyramid edge features with current-stage features and merges them using a 1×1 convolution to enhance representation and suppress noise.The design integrates boundary information at different granularities before forwarding the output to the mini-MTL module.
  • Mini Multi-Task Learning Module: The mini-MTL module jointly learns object-mask segmentation and lesion-boundary detection, using task complementarity to promote finer feature extraction.The supplied figure caption identifies two task-specific branches and an interactive attention layer.

C. Mini Multi-Task Learning Module

The mini-MTL module jointly learns object segmentation and lesion-boundary detection, using interactive attention to exchange task-specific information and refine boundary details.

  • C. Mini Multi-Task Learning Module: The mini-MTL module jointly learns object masks and lesion boundaries at each encoder stage with limited parameter overhead.It uses complementary information from semantic segmentation and edge detection.
  • C. Mini Multi-Task Learning Module: Interactive attention uses sigmoid-derived masks and reverse attention to selectively transfer useful features between the edge and segmentation branches.The same integration principle is applied symmetrically to both task branches.
  • C. Mini Multi-Task Learning Module: The gated interactive attention mechanism suppresses irrelevant information on both sender and receiver sides without adding parameters.It regulates task information spatially before aggregating the two branches.
  • C. Mini Multi-Task Learning Module: The module aggregates the two task branches after convolution and interaction to produce a richer context representation for the current encoder stage.The resulting representation is formed through concatenation and 1 × 1 convolution.
  • C. Mini Multi-Task Learning Module: Joint supervision trains edge and segmentation outputs toward their respective ground-truth masks using binary cross-entropy losses.This joint learning preserves fine details near boundaries and produces more boundary-adherent features.

D. Cross Features Fusion Module

The cross feature fusion module selectively combines complementary encoder features from different levels, retaining spatial details and semantic information while limiting redundancy.

  • D. Cross Features Fusion Module: CFF adaptively selects complementary components from feature maps at other encoder levels through an attention mechanism.It uses the differing strengths of low-level spatial detail and high-level semantic information.
  • D. Cross Features Fusion Module: CFF integrates information from different levels while avoiding excessive redundant information.The resulting contextual features preserve both details and semantic information for decoder processing.

E. Decoding and Optimization

The decoder aggregates ASPP output and encoder features stage by stage for final prediction, while joint losses train the network and incorporate boundary guidance.

  • E. Decoding and Optimization: Decoder features D_i are obtained by successively aggregating ASPP output features with encoder features from each stage.The resulting decoding features are used sequentially to produce the final segmentation prediction.
  • E. Decoding and Optimization: The network uses standard binary cross-entropy to minimize the error between decoder outputs and ground-truth masks.Training is end-to-end and includes joint losses from the multi-task module.
  • E. Decoding and Optimization: Boundary information guides final segmentation through the loss, making the network boundary-aware and refining the prediction.The balance parameters λ_i are empirically set to 1.0.

IV. EXPERIMENTS

BA-Net is evaluated across five medical-image datasets spanning multiple modalities and compared with six state-of-the-art segmentation methods under common preprocessing and metrics.

  • IV. EXPERIMENTS: Experiments cover five datasets with dermoscopy, endoscopic, X-ray, and retinal fundus images.The described ISIC-2017 dataset contains 2,000 annotated training images and 600 testing images.
  • IV. EXPERIMENTS: BA-Net is compared with FCN, U-net, MultiResUNet, AG-net, CE-Net, and Deeplabv3 using consistent preprocessing and evaluation metrics.Segmentation maps come from authors’ released code or directly provided predictions.

C. Evaluation Metric

BA-Net is evaluated with five pixel-level segmentation metrics, while Jaccard Index is used as the primary ranking criterion. Experiments use standardized augmentation, resizing, and benchmark datasets.

  • Evaluation criteria: Five metrics—Dice Similarity Coefficient, Jaccard Index, Accuracy, Sensitivity, and Specificity—are computed for evaluation.The metrics are defined from pixel-level true-positive, true-negative, false-positive, and false-negative counts.
  • Evaluation criteria: Jaccard Index primarily reflects overlap between estimated and ground-truth masks and is the main performance-ranking metric.
  • Implementation setting: All models use consistent input preprocessing, including random flips, rotations, center cropping, and image resizing.Training and testing images are uniformly resized to 256 × 256.
  • Benchmark organization: The experiments report segmentation performance on four benchmark datasets in Table I and on ISIC-2017 in Table II.The supplied table captions identify the benchmark coverage and ISIC-2017 evaluation.

E. Comparisons with the State-of-the-Art

BA-Net achieves strong segmentation results across skin-lesion, polyp, lung, and optic-disc benchmarks, with reported improvements over competing methods on several datasets. The comparisons support consistent gains across diverse medical imaging tasks.

  • Dermoscopy image dataset: BA-Net achieves the highest Dice of 88.2% and Jaccard of 81.0% for skin-lesion segmentation on ISIC-2017.Its Jaccard improves from the best competitor’s 78.5% to 81.0%, exceeding FCN and U-net by 5.8% and 4.5%.
  • Dermoscopy image dataset: BA-Net outperforms the best ISIC-2017 challenge result by 4.5% on Jaccard and exceeds MB-DCNN by 0.6% on Jaccard.The paper reports advantages over both challenge submissions and published benchmarks.
  • Endoscopic image dataset: BA-Net attains Jaccard of 86.1% on Kvasir-SEG, exceeding CE-Net and Deeplabv3 by 2.6%.It is also reported as the best-performing method on both Kvasir-SEG and CVC-ColonDB.
  • Lung segmentation: For lung segmentation in 2D X-ray images, BA-Net reaches an overall Jaccard of 92.8%, increasing 5.9% over FCN.The paper describes its performance as state of the art on most reported metrics.
  • Optic-disc segmentation: On RIM-ONE-R1 optic-disc segmentation, BA-Net achieves the best results for both single-expert evaluation and overall average evaluation.The comparison uses annotations from five independent experts.

5) Qualitative Evaluation:

Qualitative comparisons show that BA-Net produces more accurate object localization and details while suppressing similar background regions and noise. Ablation visualizations associate these improvements with the network’s three modules.

  • Qualitative comparison: BA-Net highlights objects with accurate locations and details while suppressing similar background regions and noise more thoroughly than competing methods.Other methods sometimes segment only part of lesions or respond to non-foreground information.
  • Ablation analysis: Ablation experiments on ISIC-2017 and Kvasir-SEG evaluate the effects of removing PEE, mini-MTL, and CFF modules.The ablations are reported in Table IV.
  • Ablation analysis: Removing PEE decreases Jaccard by 1.0% and 1.6% on the two datasets, respectively.The paper attributes this decline to reduced use of edge features.
  • Ablation analysis: Removing mini-MTL decreases Jaccard by 0.8% and 1.3%, weakening use of complementary boundary information.
  • Ablation analysis: CFF improves Jaccard by 1.1% and 0.7% over a model using only PEE and mini-MTL on the two tasks.The paper links this gain to integrating features across levels.

2) Effectiveness of the Interactive Attention:

Interactive attention improves the exchange of information between the edge and segmentation branches, while the complete cascaded design provides stronger segmentation than ablated variants. The conclusion extends evaluation to five public datasets and identifies future 3D segmentation as a direction.

  • Interactive attention: Removing interactive attention lowers Jaccard by 0.7% and 1.1% on the two evaluated datasets.The paper states that interaction between the edge and segmentation subnetworks improves the final representation.
  • Ablation visualization: Visualizations show the complete BA-Net better highlights objects than variants without PEE, mini-MTL, CFF, or interactive attention.The complete model avoids introducing background regions more effectively in the reported examples.
  • Module roles: PEE supplies boundary guidance, mini-MTL uses boundary-information complementarity, CFF enables cross-level interaction, and interactive attention enhances information delivery between branches.
  • Conclusion: BA-Net is evaluated on five publicly available datasets, with the reported comparative results validating its segmentation performance.The paper states that extending BA-Net to 3D medical image segmentation is future work.
Loading 2005.00966v1…