Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

X-ray Insights Unleashed: Pioneering the Enhancement of Multi-Label Long-Tail Data

Xinquan Yang, Jinheng Xie, Yawen Huang, Yuexiang Li, Huimin Huang, Hao Zheng, Xian Wu, Yefeng Zheng, Linlin Shen

TL;DR

This work addresses the long-tail problem in multi-label chest X-ray classification by proposing a normal-X-ray diffusion inpainting pipeline that augments tail-class lesions. The method trains a diffusion model on abundant normal X-rays to inpaint head-class lesion regions, then uses Grad-CAM to localize regions and an LLM-based knowledge guidance module (LKG) together with Progressive Incremental Learning (PIL) to stabilize training. Key contributions include a publicly released normal-X-ray generator, LKG to resolve disease entanglement, PIL to integrate many-tail data without catastrophic forgetting, and state-of-the-art tail-class performance on MIMIC-CXR and CheXpert. The findings offer a practical, data-efficient route to improve tail-class detection in real-world clinical datasets.

Abstract

Long-tailed pulmonary anomalies in chest radiography present formidable diagnostic challenges. Despite the recent strides in diffusion-based methods for enhancing the representation of tailed lesions, the paucity of rare lesion exemplars curtails the generative capabilities of these approaches, thereby leaving the diagnostic precision less than optimal. In this paper, we propose a novel data synthesis pipeline designed to augment tail lesions utilizing a copious supply of conventional normal X-rays. Specifically, a sufficient quantity of normal samples is amassed to train a diffusion model capable of generating normal X-ray images. This pre-trained diffusion model is subsequently utilized to inpaint the head lesions present in the diseased X-rays, thereby preserving the tail classes as augmented training data. Additionally, we propose the integration of a Large Language Model Knowledge Guidance (LKG) module alongside a Progressive Incremental Learning (PIL) strategy to stabilize the inpainting fine-tuning process. Comprehensive evaluations conducted on the public lung datasets MIMIC and CheXpert demonstrate that the proposed method sets a new benchmark in performance.

X-ray Insights Unleashed: Pioneering the Enhancement of Multi-Label Long-Tail Data

TL;DR

This work addresses the long-tail problem in multi-label chest X-ray classification by proposing a normal-X-ray diffusion inpainting pipeline that augments tail-class lesions. The method trains a diffusion model on abundant normal X-rays to inpaint head-class lesion regions, then uses Grad-CAM to localize regions and an LLM-based knowledge guidance module (LKG) together with Progressive Incremental Learning (PIL) to stabilize training. Key contributions include a publicly released normal-X-ray generator, LKG to resolve disease entanglement, PIL to integrate many-tail data without catastrophic forgetting, and state-of-the-art tail-class performance on MIMIC-CXR and CheXpert. The findings offer a practical, data-efficient route to improve tail-class detection in real-world clinical datasets.

Abstract

Long-tailed pulmonary anomalies in chest radiography present formidable diagnostic challenges. Despite the recent strides in diffusion-based methods for enhancing the representation of tailed lesions, the paucity of rare lesion exemplars curtails the generative capabilities of these approaches, thereby leaving the diagnostic precision less than optimal. In this paper, we propose a novel data synthesis pipeline designed to augment tail lesions utilizing a copious supply of conventional normal X-rays. Specifically, a sufficient quantity of normal samples is amassed to train a diffusion model capable of generating normal X-ray images. This pre-trained diffusion model is subsequently utilized to inpaint the head lesions present in the diseased X-rays, thereby preserving the tail classes as augmented training data. Additionally, we propose the integration of a Large Language Model Knowledge Guidance (LKG) module alongside a Progressive Incremental Learning (PIL) strategy to stabilize the inpainting fine-tuning process. Comprehensive evaluations conducted on the public lung datasets MIMIC and CheXpert demonstrate that the proposed method sets a new benchmark in performance.
Paper Structure (16 sections, 3 equations, 5 figures, 6 tables)

This paper contains 16 sections, 3 equations, 5 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Lung Disease Recognition
  4. Long Tail Data Augmentation
  5. Method
  6. Data Collection and Model Pre-training
  7. DiT Training on Normal X-rays
  8. Lung Lesion Classifier Training on Abnormal X-rays
  9. Tail-class Lesion Data Generation
  10. LLM Knowledge Guidance
  11. Progressive Incremental Finetuning
  12. Experiments
  13. Datasets
  14. Implementation Details
  15. Performance Analysis
  16. ...and 1 more sections

Figures (5)

  • Figure 1: Comparison between previous Chest X-ray data augmentation and the proposed method. (a) Previous method based on diseased data for generation. (b) Our method based on normal data for inpainting.
  • Figure 2: (a) The data distribution in each public dataset. (b) Diagram of lung disease entanglement. Each point in the figure is the center of the lesion annotation in the VinDr-CXR nguyen2022vindr dataset.
  • Figure 3: Overview of our proposed framework.
  • Figure 4: Visualization of the inpainting result generated by the normal X-ray diffusion model.
  • Figure 5: The architecture of DiT variant.