Joint Medical Image Enhancement and Segmentation with Diffusion-based Symbiotic Information Interaction

Technical Summary: DiSIINet – Joint Medical Image Enhancement and Segmentation with Diffusion-based Symbiotic Information Interaction

Problem Statement
Medical imaging modalities such as MRI, CT, and ultrasound frequently suffer from quality degradation due to cost constraints, resulting in low resolution, noise, blurriness, and poor lighting. These limitations complicate the visualization of critical anatomical structures and lesions, potentially leading to misdiagnosis. Traditional approaches typically treat image enhancement as a separate preprocessing step before segmentation, failing to leverage the potential synergy between the two tasks. Furthermore, existing generative methods, including GANs and early diffusion models, often struggle to balance perceptual quality with fidelity, or rely on unidirectional information flow (e.g., using segmentation to guide enhancement) that does not fully exploit mutual benefits. Recent text-guided diffusion approaches face limitations in semantic accuracy, such as overlooking key objects or creating ambiguity in super-resolution tasks.

Methodology
The authors propose DiSIINet (Diffusion-based Symbiotic Information Interaction Network), a unified framework that jointly performs medical image enhancement and segmentation. The architecture is built upon Denoising Diffusion Implicit Models (DDIM) and consists of two primary branches:

1. DiEnh (Diffusion-based Enhancement): A branch dedicated to restoring low-quality (LQ) images to high-quality (HQ) images.
2. DiSeg (Diffusion-based Segmentation): A branch dedicated to generating segmentation masks from LQ images.

Key Mechanism: Symbiotic Information Interaction (SII)
The core innovation is the Symbiotic Information Interaction (SII) module, which facilitates dynamic, bidirectional information exchange between the DiEnh and DiSeg branches during the reverse diffusion process. Unlike unidirectional cascaded methods, DiSIINet allows the tasks to iteratively improve each other.

• Architecture: The SII module comprises an Enh-Controller and a Seg-Controller.
• Operation: At each reverse diffusion step $t$, the module computes cross-attention between the latent representations of both branches.
  • The Enh-Controller treats the segmentation latent ($Z^{seg}_t$) as the query and the enhancement latent ($Z^{enh}_t$) as the key and value matrices to generate a guided feature ($F^{(t)}_{oi}$) for the enhancement branch.
  • The Seg-Controller treats the enhancement latent ($Z^{enh}_t$) as the query and the segmentation latent ($Z^{seg}_t$) as the key and value matrices to generate a guided feature ($F^{(t)}_{os}$) for the segmentation branch.
• Training: The model is trained end-to-end using a joint loss function that combines DDIM noise prediction losses for both branches, a restoration loss (MSE) for the enhanced image, and a weighted binary cross-entropy loss for the segmentation mask.

Key Contributions

1. Dual-Branch Diffusion Architecture: DiSIINet is presented as the first framework to utilize a dual-branch diffusion architecture where enhancement and segmentation mutually reinforce each other through continuous information exchange, rather than operating sequentially or independently.
2. Symbiotic Information Interaction (SII) Module: The authors designed a novel module that enables bidirectional feature flow via cross-attention during the reverse diffusion process, allowing the segmentation branch to guide enhancement and vice versa.
3. Comprehensive Validation: The method was validated across diverse mainstream medical imaging modalities (MRI, CT, and Ultrasound) and multiple datasets (ACDC, KiTS19, TN3K), demonstrating consistent performance improvements over state-of-the-art sequential and independent approaches.

Experimental Results
Extensive experiments were conducted on three datasets:

• ACDC (MRI): DiSIINet achieved a PSNR of 32.15 and SSIM of 0.932 for enhancement, and an avg-mIoU of 84.81 and avg-Dice of 91.78 for segmentation, outperforming methods like MedSegDiff-V2 and UNet variants.
• KiTS19 (CT): The model achieved a PSNR of 34.68 and SSIM of 0.961 for enhancement, with segmentation metrics of 75.25 mIoU and 85.87 Dice.
• TN3K (Ultrasound): DiSIINet reached a PSNR of 34.26 and SSIM of 0.942, with segmentation metrics of 75.18 mIoU and 85.82 Dice.

Qualitative results indicate that DiSIINet preserves fine structural details (e.g., ventricular regions, kidney tumors, thyroid nodules) more effectively than competing methods, which often exhibit blurring or loss of detail. Ablation studies confirmed that removing either the Enh-Controller or Seg-Controller significantly degrades performance, validating the necessity of the symbiotic interaction.

Significance and Claims
The paper claims that DiSIINet addresses the gap in deep feature interaction between enhancement and segmentation tasks within the generative framework of diffusion models. By establishing a symbiotic relationship where both tasks benefit from the other's semantic and structural information, the framework achieves superior reconstruction quality and segmentation accuracy compared to unidirectional or independent methods. The authors acknowledge that these performance gains come at the cost of increased computational complexity and suggest that future work will focus on model distillation and adaptive mechanisms to reduce denoising steps without compromising the symbiotic benefits.