Lesion region inpainting: an approach for pseudo-healthy image synthesis in intracranial infection imaging

Xiaojuan Liu ^1,2, Cong Xiang ¹, Libin Lan ³

⁰College of Artificial Intelligence, Chongqing University of Technology, Chongqing, China.

Frontiers in Microbiology +

|

September 3, 2024

View abstract on PubMed

Summary

This study introduces a novel three-stage method for generating realistic pseudo-healthy medical images, crucial for disease diagnosis and data augmentation. The approach effectively preserves patient identity and reconstructs extensive lesion areas, significantly improving image quality.

Area Of Science

Medical Imaging
Artificial Intelligence
Computer Vision

Background

Pseudo-healthy image synthesis is vital for data augmentation and disease diagnosis.
Existing Generative Adversarial Network (GAN) methods struggle with heterogeneous intracranial infections, losing healthy information and identity.
Extensive lesions in pathological images often result in generated pseudo-healthy images lacking distinct structures.

Purpose Of The Study

To develop a robust method for synthesizing pseudo-healthy medical images that preserves subject identity and accurately reconstructs lesion areas.
To overcome the limitations of current GANs in handling diverse pathologies and extensive lesions.
To improve the quality and realism of generated pseudo-healthy images for clinical applications.

Main Methods

A three-stage approach: lesion localization using a Segmentor, healthy outline construction via Vague-Filler for inpainting, and realistic image synthesis using a GAN with a contextual residual attention module.
The Segmentor identifies pathological regions, distinguishing them from healthy tissue.
Vague-Filler creates a healthy outline to prevent structural loss, especially in cases of large lesions.

Main Results

The proposed method achieved a healthiness score of 0.957 on the BraTS2021 dataset across various modalities.
Generated pseudo-healthy images demonstrated superior visual quality and enhanced repair of large lesion areas compared to existing methods.
The model showed effectiveness in partially reconstructing other organs on the COVID-19-20 dataset.

Conclusions

The three-stage method successfully generates high-quality pseudo-healthy images with excellent identity preservation and lesion reconstruction capabilities.
This approach significantly advances the field of medical image synthesis, offering improved diagnostic support and data augmentation.
The model's versatility is highlighted by its performance on different datasets and its ability to reconstruct diverse anatomical structures.

Keywords:

contextual residual attention module lesion inpainting for pseudo-healthy synthesis data augmentation generative adversarial networks intracranial infection pseudo-healthy image synthesis