Related Experiment Video
Updated: Aug 12, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.9K
Super-resolution and segmentation deep learning for breast cancer histopathology image analysis
Aniwat Juhong1,2, Bo Li1,2, Cheng-You Yao2,3
1Department of Electrical and Computer Engineering, Michigan State University, East Lansing, MI 48823, USA.
Biomedical Optics Express
|January 26, 2023
Summary
Custom deep learning models enhance low-resolution histopathology images for improved cancer diagnosis. This approach overcomes limitations of high-performance microscopes, enabling better cell and nuclei characterization in resource-constrained settings.
Area of Science:
- Medical imaging
- Computational pathology
- Artificial intelligence in healthcare
Background:
- High-resolution microscopy is essential for accurate histopathological analysis but generates large image files, posing storage and transfer challenges.
- Current image compression techniques often degrade resolution, hindering detailed cancer diagnosis.
- Access to advanced microscopes and whole slide imaging systems is limited in many remote or resource-constrained settings.
Purpose of the Study:
- To develop custom Convolutional Neural Networks (CNNs) for super-resolution image enhancement of low-resolution histopathology images.
- To enable accurate characterization and segmentation of cells and nuclei from Hematoxylin and Eosin (H&E) stained breast cancer images.
- To facilitate cancer diagnosis in low-resource settings by overcoming limitations of traditional microscopy.
Main Methods:
- Utilized a super-resolution generative adversarial network based on aggregated residual transformation (SRGAN-ResNeXt) for image enhancement.
- Employed a separate custom CNN for image segmentation of cells and nuclei from the enhanced high-resolution images.
- Investigated jointly trained SRGAN-ResNeXt and Inception U-net models using transfer learning for improved performance.
Main Results:
- Achieved significant image quality enhancement with Peak Signal-to-Noise Ratio > 30 dB and Structural Similarity > 0.93.
- Demonstrated superior performance compared to bicubic interpolation and standard SRGAN methods.
- Obtained high accuracy in image segmentation with an average Intersection over Union of 0.869 and Dice Similarity Coefficient of 0.893.
- Jointly trained models showed progressively improved and promising results.
Conclusions:
- Custom CNNs, particularly the SRGAN-ResNeXt and Inception U-net models, effectively enhance low-resolution histopathology images.
- The developed models facilitate accurate cell and nuclei segmentation, crucial for cancer diagnosis.
- This AI-driven approach offers a viable solution for high-resolution image acquisition and analysis in resource-limited environments, democratizing access to advanced diagnostic tools.
