Automatic Image Generation for Breast Cancer Diagnosis
Author Information
Author(s): Saad Afaf, Ghatwary Noha, Gasser Safa M., ElMahallawy Mohamed S.
Primary Institution: Arab Academy for Science, Technology & Maritime Transport
Hypothesis
Can a deep learning model effectively generate high-quality IHC-stained images from H&E-stained images to improve breast cancer diagnosis?
Conclusion
The proposed IHC-GAN model successfully generates high-quality IHC images from H&E images, significantly reducing costs and improving efficiency in breast cancer diagnostics.
Supporting Evidence
- The IHC-GAN model achieved a FID score of 0.0927, indicating high fidelity in image generation.
- It demonstrated a PSNR of 22.87, reflecting superior image detail preservation.
- The model's SSIM score was 0.3735, showing good structural similarity to reference images.
- The proposed approach resulted in an 88% reduction in FID compared to existing GAN models.
- The model was trained on a comprehensive dataset of 4,870 image pairs.
Takeaway
This study created a smart computer program that helps doctors see breast cancer better by turning one type of image into another, making it easier and cheaper to diagnose.
Methodology
The study used a deep learning model called IHC-GAN, trained on a dataset of 4,870 image pairs, to convert H&E stained images into IHC stained images.
Potential Biases
The model's reliance on the MobileNetV3 classifier introduces risks of misclassification affecting the staining accuracy.
Limitations
The model's performance may be affected by the quality of input data, and downscaling images led to some loss of detail.
Participant Demographics
The dataset included various levels of HER2 expression (0, 1+, 2+, 3+).
Digital Object Identifier (DOI)
Want to read the original?
Access the complete publication on the publisher's website