DiffSeg: Diffusion-Based Weakly Supervised Segmentation for Fibrosis

Publication & Presentation

This project was published at IEEE ISBI 2025 and featured as an invited presentation in April 2025, Houston, TX, USA.

Introduction

Fibrotic Lung Disease (FLD) is a life-threatening condition characterised by progressive lung scarring, responsible for 1% of UK deaths in 2021. However, fibrosis appears in irregular and diffuse patterns with poorly defined boundaries, making manual segmentation highly variable and labor-intensive.

To address this, we propose an automated segmentation pipeline that avoids pixel-level annotations and instead relies on image-level binary labels, reducing human annotation effort while maintaining accuracy.

Method

We introduce DiffSeg, a novel diffusion-based weakly supervised semantic segmentation framework. Key innovations include:

• Image-level supervision: Only binary (yes/no) fibrosis labels are used.
• Controllable image generation: Using a diffusion model, healthy CT slices are transformed into synthetic images with controllable fibrosis severity.
• Frozen classifier and refinement pipeline: Paired healthy and fibrosis-injected images help guide a classifier to generate accurate pseudo masks.

The pipeline synthesizes diverse training data and leverages pseudo-label refinement for robust segmentation, even in complex pathological patterns.

Results

DiffSeg achieves state-of-the-art performance with a Dice score of 61.75% (IQR: 52.37–70.02%), outperforming existing methods:

Visual comparisons show that DiffSeg closely resembles ground truth masks with finer details and minimal noise, demonstrating strong generalizability without relying on pixel-wise labels.

Conclusion

DiffSeg is a robust and efficient framework for weakly supervised segmentation in medical imaging. By leveraging generative diffusion models, it:

• Reduces the need for expensive expert annotations,
• Achieves competitive segmentation performance, and
• Enables fine-grained analysis of complex fibrosis patterns.

This work opens new directions for scalable, label-efficient medical AI systems, particularly where expert annotations are difficult or costly to obtain.

📄 Full Poster

Project Links:
📄 Paper on arXiv
📫 Contact: zhilinggavin@gmail.com
🔗 LinkedIn