Frequency Domain Unlocks New Perspectives for Abdominal Medical Image Segmentation

TL;DR

The paper tackles the problem of accurate segmentation in low-contrast abdominal medical images, where foreground regions are difficult to distinguish from surrounding tissue. It introduces the Foreground-Aware Spectrum Segmentation (FASS) framework, which combines a Foreground-Aware module, a Feature-Level Frequency Enhancement (FLFE) module based on wavelet decomposition with cross-attention, and an Edge Constraint (EC) module to enforce boundary continuity. Training relies on a distribution-divergence loss $\\ell_{KL}$, a Gaussian heating schedule $\\lambda(t)$, and a boundary-coherence loss $\\mathcal{L}_{EC}$ that aggregates $\\mathcal{L}_{match}$ and $\\mathcal{L}_{cont}$ into $\\mathcal{L}_{EC}$, with the total loss defined as $\\mathcal{L}_{total}=\\mathcal{L}_{sup}+\\lambda(t)(\\mathcal{L}_D+\\mathcal{L}_{EC})$ and $\\mathcal{L}_{sup}=\\tfrac{1}{2}(\\mathcal{L}_{Dice}+\\mathcal{L}_{CE})$. Experimental results on MSD Pancreas, NIH, and LiMT show that FASS achieves state-of-the-art performance across Dice, Jaccard, 95HD, and ASD, with improved boundary integrity and robustness in fine structures. Overall, FASS offers a principled, frequency-aware approach to enhance foreground localization and boundary fidelity in challenging low-contrast clinical scenarios, with potential for broader adoption in diverse medical imaging tasks.

Abstract

Accurate segmentation of tumors and adjacent normal tissues in medical images is essential for surgical planning and tumor staging. Although foundation models generally perform well in segmentation tasks, they often struggle to focus on foreground areas in complex, low-contrast backgrounds, where some malignant tumors closely resemble normal organs, complicating contextual differentiation. To address these challenges, we propose the Foreground-Aware Spectrum Segmentation (FASS) framework. First, we introduce a foreground-aware module to amplify the distinction between background and the entire volume space, allowing the model to concentrate more effectively on target areas. Next, a feature-level frequency enhancement module, based on wavelet transform, extracts discriminative high-frequency features to enhance boundary recognition and detail perception. Eventually, we introduce an edge constraint module to preserve geometric continuity in segmentation boundaries. Extensive experiments on multiple medical datasets demonstrate superior performance across all metrics, validating the effectiveness of our framework, particularly in robustness under complex conditions and fine structure recognition. Our framework significantly enhances segmentation of low-contrast images, paving the way for applications in more diverse and complex medical imaging scenarios.

Figures (7)

• Figure 1: Comparison between our FASS framework and previous automatic segmentation methods. (a) Segmentation process of low-contrast images by previous methods, which face challenges such as insufficient target perception, undiscriminating features, and broken edges. (b) Segmentation process of the FASS method. Our FASS framework employs adversarial training between the full image and background feature distribution to achieve focused attention on the foreground. Discriminative features are then enhanced in the frequency domain, and boundary integrity and continuity are strengthened through the edge constraint module.
• Figure 2: Overview of the proposed FASS framework. The framework consists of a foreground-aware (FA) module (Sec. \ref{['sec3.1']}), a feature-level frequency enhancement (FLFE) module (Sec. \ref{['sec3.2']}), and an edge constraint (EC) module (Sec. \ref{['sec3.3']}). Initially, sampled patches and background patches are fed into the encoder for feature extraction. The feature differences between encoder outputs are computed and maximized, with the FLFE module enhancing features during the encoding phase. Finally, the EC module refines the edges of the decoder output for optimized segmentation results.
• Figure 3: Qualitative segmentation examples of our framework compared to competing approaches across three datasets demonstrate that our framework significantly enhances segmentation accuracy and integrity, especially in capturing complex tumor shapes and detecting small tumors.
• Figure 4: Visual analysis of encoder feature extraction. With the introduction of the FA module, our method effectively focuses on foreground areas, filtering out the complex background information.
• Figure 5: Visualization comparison of high-frequency features: (a) shows the original image, (b) displays the feature map extracted after the separable convolution operation, and (c)-(f) present the high-frequency components in the vertical, horizontal, and diagonal directions, respectively. (g) illustrates the feature map after integrating each component through the cross-attention mechanism.