Prompt2SegCXR:Prompt to Segment All Organs and Diseases in Chest X-rays

TL;DR

Prompt2SegCXR is a novel lightweight model that enables flexible, prompt-based multi-organ and multi-disease segmentation in Chest X-rays, addressing limitations of traditional models and enhancing medical image analysis.

Contribution

The paper introduces a new prompt-based segmentation dataset with expert doodle prompts and a lightweight model with multi-stage feature fusion for improved accuracy.

Findings

Effective multi-organ and multi-disease segmentation in Chest X-rays.

Model outperforms existing prompt-based segmentation models.

Provides a reliable, flexible tool for medical image analysis.

Abstract

Image segmentation plays a vital role in the medical field by isolating organs or regions of interest from surrounding areas. Traditionally, segmentation models are trained on a specific organ or a disease, limiting their ability to handle other organs and diseases. At present, few advanced models can perform multi-organ or multi-disease segmentation, offering greater flexibility. Also, recently, prompt-based image segmentation has gained attention as a more flexible approach. It allows models to segment areas based on user-provided prompts. Despite these advances, there has been no dedicated work on prompt-based interactive multi-organ and multi-disease segmentation, especially for Chest X-rays. This work presents two main contributions: first, generating doodle prompts by medical experts of a collection of datasets from multiple sources with 23 classes, including 6 organs and 17 diseases, specifically designed for prompt-based Chest X-ray segmentation. Second, we introduce Prompt2SegCXR, a lightweight model for accurately segmenting multiple organs and diseases from Chest X-rays. The model incorporates multi-stage feature fusion, enabling it to combine features from various network layers for better spatial and semantic understanding, enhancing segmentation accuracy. Compared to existing pre-trained models for prompt-based image segmentation, our model scores well, providing a reliable solution for segmenting Chest X-rays based on user prompts.