PaperlayerView exact revision on arXiv

Source-linked AI summary

ConDSeg: A General Medical Image Segmentation Framework via Contrast-Driven Feature Enhancement

Mengqi Lei, Haochen Wu, Xinhua Lv, Xin Wang

arXiv:2412.08345v1eess.IVcs.CV

TL;DR

Medical-image segmentation is challenged by ambiguous boundaries and co-occurrence that can mislead models. ConDSeg combines robust encoder training, feature decoupling and contrast-guided fusion, and size-specific decoding; it reports state-of-the-art performance across five datasets and three modalities.

  • Problem

    Soft boundaries, poor illumination, and low contrast obscure foreground-background distinctions, while co-occurrence can lead models to learn misleading patterns.

    Source: Introduction · p. 1

  • Method

    ConDSeg uses consistency-based encoder training, decouples features into foreground, background, and uncertainty, then applies contrast-guided fusion and size-specific decoding.

    Source: Introduction · p. 2, Network Architecture · p. 3

  • Results

    ConDSeg achieves state-of-the-art performance across five medical-image datasets covering three modalities.

    Source: Introduction · p. 2

  • Takeaways & Limitations

    The framework addresses boundary ambiguity and co-occurrence across the evaluated medical-image segmentation tasks.

    Source: Introduction · p. 2

Abstract

from arXiv · show

Medical image segmentation plays an important role in clinical decision making, treatment planning, and disease tracking. However, it still faces two major challenges. On the one hand, there is often a ``soft boundary'' between foreground and background in medical images, with poor illumination and low contrast further reducing the distinguishability of foreground and background within the image. On the other hand, co-occurrence phenomena are widespread in medical images, and learning these features is misleading to the model's judgment. To address these challenges, we propose a general framework called Contrast-Driven Medical Image Segmentation (ConDSeg). First, we develop a contrastive training strategy called Consistency Reinforcement. It is designed to improve the encoder's robustness in various illumination and contrast scenarios, enabling the model to extract high-quality features even in adverse environments. Second, we introduce a Semantic Information Decoupling module, which is able to decouple features from the encoder into foreground, background, and uncertainty regions, gradually acquiring the ability to reduce uncertainty during training. The Contrast-Driven Feature Aggregation module then contrasts the foreground and background features to guide multi-level feature fusion and key feature enhancement, further distinguishing the entities to be segmented. We also propose a Size-Aware Decoder to solve the scale singularity of the decoder. It accurately locate entities of different sizes in the image, thus avoiding erroneous learning of co-occurrence features. Extensive experiments on five medical image datasets across three scenarios demonstrate the state-of-the-art performance of our method, proving its advanced nature and general applicability to various medical image segmentation scenarios. Our released code is available at \url{https://github.com/Mengqi-Lei/ConDSeg}.

Introduction

ConDSeg targets ambiguous boundaries and misleading co-occurrence in medical images with a framework combining robust feature learning, foreground-background contrast, and size-specific decoding. The paper reports state-of-the-art performance across five datasets and three modalities.

Introduction

  • Medical-image segmentation is difficult because soft boundaries, poor illumination, and low contrast obscure foreground-background distinctions, while co-occurrence can mislead models.Source: Introduction · p. 1
  • Consistency Reinforcement trains the encoder on original and strongly augmented images, maximizing agreement between predicted masks to improve robustness under varied lighting and color conditions.Source: Introduction · p. 2
  • Semantic Information Decoupling separates encoder features into foreground, background, and uncertainty regions, with training designed to reduce uncertainty and improve mask precision.Source: Introduction · p. 2
  • ConDSeg achieves state-of-the-art performance across five medical-image datasets covering three modalities.Source: Introduction · p. 2
  • Contrast-Driven Feature Aggregation uses foreground-background contrast to guide multilevel feature fusion, while the Size-Aware Decoder predicts entities of different sizes separately to address co-occurrence interference.Source: Introduction · p. 2

Related Works

Prior medical-image segmentation work has improved encoder-decoder architectures and addressed boundary ambiguity with boundary-focused modules. The section also identifies co-occurrence as a source of misleading contextual reliance and inaccurate predictions.

Related Works

  • Encoder-decoder methods have advanced through skip connections, multiscale context, edge features, auxiliary text supervision, and Transformer encoders.Source: Related Works · p. 2
  • Boundary-focused methods add boundary prediction, boundary-sensitive losses, or edge-guided features, but the review says they do not fundamentally reduce uncertainty in ambiguous regions.Source: Related Works · p. 2
  • Co-occurring structures can encourage models to rely on contextual associations, so a solitary polyp may be predicted as multiple polyps.Source: Related Works · pp. 2–3

Methodology

ConDSeg uses a two-stage encoder-decoder pipeline: consistency-based encoder pretraining is followed by feature decoupling, contrast-guided fusion, and size-specific decoding.

Network Architecture

  • The first stage uses Consistency Reinforcement to preliminarily train the encoder for robustness to weak illumination and low contrast; the second fine-tunes it at a lower learning rate.Source: Network Architecture · p. 3
  • The ResNet-50 encoder extracts four feature levels, and SID decouples the deepest feature into foreground, background, and uncertainty information.Source: Network Architecture · p. 3
  • CDFA fuses encoder feature levels under foreground-background guidance, after which level-specific decoders locate entities by size and fuse their outputs into the final mask.Source: Network Architecture · p. 3

Improving Boundary Distinction

Consistency Reinforcement trains the encoder to produce consistent masks for original and strongly augmented images, supporting feature extraction across adverse imaging conditions. Semantic Information Decoupling separates encoder features into foreground, background, and uncertainty representations, with training constraints that reduce uncertainty.

Consistency Reinforcement

  • Consistency Reinforcement compares predictions from original and strongly augmented images, using mask losses and a consistency loss to improve encoder robustness across lighting and contrast changes.The consistency loss alternately binarizes one prediction as a reference for BCE against the other, avoiding the numerical instability associated with KL and JS divergence for probabilities near 0 or 1.Source: Consistency Reinforcement · pp. 3–4

Semantic Imformation Decoupling

  • Semantic Information Decoupling maps encoder features into foreground, background, and uncertainty branches, whose predictions are constrained toward complementary pixel categories.The foreground and background predictions are optimized toward the ground truth and its negation, while the three masks are constrained to be complementary.Source: Semantic Imformation Decoupling · p. 4
  • Area-ratio penalty terms weight foreground and background losses to give smaller entities more attention during training.The penalties are calculated from prediction-area ratios and are intended to improve loss stability.Source: Semantic Imformation Decoupling · p. 4
  • Complementarity and branch-specific accuracy constraints encourage the model to reduce uncertainty regions while strengthening foreground and background features.The complementarity loss assigns each pixel exclusively to foreground, background, or uncertainty.Source: Semantic Imformation Decoupling · pp. 4–5

Overcoming Co-occurrence

ConDSeg addresses co-occurrence errors through contrast-guided feature aggregation and size-specific decoding. CDFA uses foreground and background features to guide multilevel fusion, while the decoder assigns different entity scales to separate paths.

Overcoming Co-occurrence

  • ConDSeg avoids TGANet’s tendency to predict multiple polyps when a single polyp appears, using CDFA and the Size-Aware Decoder to address co-occurrence.The comparison uses Grad-CAM visualizations on Kvasir-SEG examples.Source: Overcoming Co-occurrence · p. 5

Contrast-Driven Feature Aggregation

  • CDFA uses decoupled foreground and background features to guide multilevel feature fusion and enhance key features for distinguishing entities from complex backgrounds.Local attention weights derived from foreground and background features aggregate neighboring features within a K × K window.Source: Contrast-Driven Feature Aggregation · p. 5

Size-Aware Decoder

  • The Size-Aware Decoder assigns smaller entities to shallow features and larger entities to deeper features, then fuses predictions from three size-specific decoders.The shallow features carry more fine-grained information, while deeper features contain more global information and higher-level semantics.Source: Size-Aware Decoder · p. 6

Overall Optimization and Training Process

ConDSeg uses two training stages: the first enhances encoder feature extraction and robustness, and the second optimizes the full model with a lower learning rate for the encoder.

Overall Optimization and Training Process

Experiments

ConDSeg is evaluated on five public datasets spanning three medical image modalities and is reported to achieve the best segmentation performance on all five. Training-strategy and module ablations also favor the complete two-stage framework with the proposed modules.

Experimental Setup

  • Experiments cover five public datasets across three modalities, use 256 × 256 images, and evaluate mIoU, mDSC, Recall, and Precision.The default encoder is ResNet-50; the study also reports that other backbones can be used.Source: Experimental Setup · p. 6

Comparison with Other State-of-the-Art

  • ConDSeg achieves the best segmentation performance on all five evaluated datasets among the compared state-of-the-art methods.Source: Comparison with Other State-of-the-Art · p. 7
  • On Kvasir-SEG, the complete two-stage ConDSeg achieves the fastest convergence and best performance, while one-stage training reaches an advanced level.Figure 5 compares training convergence curves; one-stage denotes training from scratch, and two-stage denotes training with CR-pretrained weights.Source: Comparison with Other State-of-the-Art · p. 7, Figure 5 · p. 7

Training Strategy

  • Ablations report that CR improves encoder feature extraction and robustness, and that two-stage training with CR in the first stage achieves the best results.The five settings compare direct training, CR, and two-stage variants with or without CR in the first stage.Source: Training Strategy · p. 7, Table 3 · p. 7

Proposed Modules

  • On GlaS, adding SID and CDFA incrementally improves mIoU and mDSC, with both metrics optimal when the SA-Decoder completes the architecture.The ablation baseline removes SID, replaces CDFA with ordinary convolutions, and uses a single-scale decoder.Source: Proposed Modules · p. 7, Table 4 · p. 7

Conclusion

ConDSeg addresses soft boundaries and co-occurrence through contrast-driven feature enhancement and a decoder designed for entities of different sizes. It achieves state-of-the-art performance across five datasets and three medical-imaging modalities.

Conclusion

  • ConDSeg targets soft boundaries and co-occurrence phenomena that challenge medical image segmentation.The framework is presented as a general medical image segmentation approach.Source: Conclusion · p. 7
  • Consistency Reinforcement trains the encoder for robustness to adverse illumination and low contrast, while Semantic Information Decoupling separates foreground, background, and uncertainty features.The decoupling module is intended to reduce uncertainty during training.Source: Conclusion · p. 7
  • Contrast-Driven Feature Aggregation uses foreground-background contrast to guide multi-level feature fusion and key-feature enhancement.The Size-Aware Decoder localizes entities of different sizes.Source: Conclusion · p. 7
  • ConDSeg achieves state-of-the-art performance across five medical image datasets spanning three modalities.The authors present this result as evidence of the framework's general applicability.Source: Conclusion · p. 7

Supplementary Material

The supplementary material reports dataset protocols and additional tests of thresholds, training choices, module settings, speed, encoders, and a 3D adaptation. It also includes a feature visualization for SID.

Dataset

  • Experiments cover five datasets across three modalities, using stated train-validation or train-validation-test splits for each dataset.Kvasir-SEG uses an 880/120 split; Kvasir-Sessile 156/20/20; GlaS 85/80; ISIC-2016 900/379; and ISIC-2017 2000/150/600.Source: Dataset · p. 10, Table 5 · p. 11

Supplementary Ablation Experiments

Supplementary Visualizations

Supplementary visualizations compare segmentation predictions across modalities and provide structural diagrams for the auxiliary head, SA-Decoder, and ablation baseline.

Supplementary Visualizations

  • Across three modalities, ConDSeg's visual predictions are reported as closest to ground truth and significantly better than the other shown methods.Figure 9 presents the segmentation-result comparison across datasets with different modalities.Source: Supplementary Visualizations · p. 11, Figure 9 · p. 12

Supplementary Module Structure Diagrams

Related papers

Loading 2412.08345v1…