Related Experiment Video
Updated: Aug 9, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.9K
Improving the Segmentation Accuracy of Ovarian-Tumor Ultrasound Images Using Image Inpainting.
Lijiang Chen1, Changkun Qiao1, Meijing Wu2
1School of Electronic and Information Engineering, Beihang University, No. 37 Xueyuan Road, Haidian District, Beijing 100191, China.
Bioengineering (Basel, Switzerland)
|February 25, 2023
Summary
This study introduces a novel mask-guided generative adversarial network (MGGAN) to remove symbols from 2D ovarian tumor ultrasound images, improving diagnostic accuracy. The MGGAN effectively cleans images without needing original clean versions, enhancing lesion segmentation and classification for better AI-driven diagnosis.
Area of Science:
- Medical Imaging
- Artificial Intelligence in Healthcare
- Computer Vision
Background:
- Diagnostic accuracy of 2D ovarian tumor ultrasound images is compromised by artificial symbols (e.g., fingers, crosses) present in clinical images.
- These symbols obscure lesion boundaries, hindering feature extraction and negatively impacting AI-based lesion classification and segmentation accuracy.
- Existing image inpainting techniques are explored for noise and object removal in medical imaging.
Purpose of the Study:
- To develop and present a novel framework, the mask-guided generative adversarial network (MGGAN), for removing artificial symbols from 2D ovarian tumor ultrasound images.
- To enhance the quality of ultrasound images for improved AI-driven diagnostic accuracy.
- To achieve pixel-level inpainting of distorted regions without requiring original clean images.
Main Methods:
- Construction of a 2D ovarian-tumor ultrasound image inpainting dataset by annotating symbols within the MMOTU dataset.
- Implementation of a mask-guided generative adversarial network (MGGAN) incorporating an attention mechanism in the generator.
- Integration of fast Fourier convolutions (FFCs) and residual networks to enhance the global perceptual field for high-resolution image processing.
Main Results:
- The MGGAN demonstrated high performance in corrupted regions by focusing on valid information and disregarding symbols, leading to more realistic lesion boundaries.
- The model achieved superior objective and subjective evaluation results compared to other models in a single stage.
- Significant improvements in segmentation accuracy were observed: Unet model accuracy increased from 71.51% to 76.06%, and PSPnet model accuracy increased from 61.13% to 66.65%.
Conclusions:
- The MGGAN effectively removes artificial symbols from 2D ovarian tumor ultrasound images, enabling pixel-level inpainting without clean reference images.
- The proposed method significantly enhances the accuracy of computerized ovarian tumor diagnosis by improving image quality for AI analysis.
- The MGGAN framework shows promise for clinical application, offering a robust solution for preprocessing ultrasound images to improve diagnostic outcomes.
