DiSIINet uses diffusion to jointly enhance and segment medical images

Researchers from academia have introduced DiSIINet (Diffusion-based Symbiotic Information Interaction Network), a unified framework that combines medical image enhancement and segmentation in a single diffusion process. The model is built on Denoising Diffusion Implicit Models (DDIM) and features two parallel branches — one for enhancement and one for segmentation — that exchange information iteratively through a novel Symbiotic Information Interaction (SII) module. This module uses cross-attention during the reverse diffusion process to align and share features between tasks, enabling each branch to guide the other toward better results.

The key innovation is that enhancement and segmentation are no longer treated as sequential preprocessing steps but as mutually reinforcing objectives. Traditional pipelines first enhance an image using super-resolution or denoising methods, then run a separate segmentation model — which often loses information between steps. DiSIINet instead allows the segmentation branch to provide anatomical context that informs the enhancement, and the enhancement branch to deliver cleaner representations that improve segmentation accuracy. The DDIM backbone ensures high-quality output with deterministic and efficient sampling.

Experiments were conducted on multi-modal medical datasets including MRI, CT, and ultrasound — modalities notorious for low resolution and noise due to cost or acquisition constraints. DiSIINet achieved significant performance improvements over both sequential and independent baselines across all three modalities. The results demonstrate that the symbiotic approach not only boosts segmentation accuracy but also produces higher-quality enhanced images, which could aid radiologists in diagnosis. The code has been released, and the paper was accepted at IJCAI 2026.