Enhancing Weakly Supervised Semantic Segmentation for Fibrosis via Controllable Image Generation
Developed AI tools integrating Generative AI for precise pattern detection and segmentation, applied to medical image analysis and automation
IEEE ISBI 2025
By Gavin (Zhiling) Yue, Yingying Fang, et al.
Figure 1. A Framework overview. (a) Pre-training of a diffusion-based autoencoder and classifier. (b) Slice-injected fibrosis generation workflow with trained models.
đź”— View Full Project
Abstract
Fibrotic Lung Disease (FLD) is a progressive and fatal condition characterized by lung stiffening and scarring, leading to respiratory decline. While high-resolution computed tomography (HRCT) is the gold standard for FLD assessment, fibrosis often appears as diffuse and irregular patterns with indistinct boundaries—posing challenges for consistent manual annotation and introducing high inter-observer variability.
To address this, we propose DiffSeg, a novel weakly supervised semantic segmentation (WSSS) framework that utilizes image-level labels to generate pixel-level fibrosis segmentations. By removing the need for pixel-level annotations, DiffSeg significantly lowers annotation cost while preserving high accuracy.
At the core of DiffSeg is a diffusion-based generative model that transforms healthy HRCT slices into fibrosis-injected counterparts, simultaneously producing paired fibrosis masks. This synthetic pairing enables more accurate and consistent pseudo-label generation than existing WSSS techniques, achieving stronger performance and improving robustness.