Virtual Immunohistochemistry for Breast Cancer Biomarker Prediction From H&E-Stained Images Using Generative Network

Immunohistochemistry (IHC) is essential in diagnostic pathology but is often constrained by cost, time, and limited tissue availability. Virtual IHC staining, which predicts IHC stains from standard hematoxylin and eosin (H&E) images, presents a promising alternative. This study introduces a novel Conditional Generative Adversarial Network (cGAN) architecture based on a U-Net with depthwise separable convolutions to enhance the accuracy and efficiency of virtual IHC staining. This architectural refinement improves computational efficiency while preserving high image quality. We trained and evaluated our model using the BCI and MIST datasets and compared its performance against established image-to-image translation techniques, including Pix2Pix, CycleGAN, and a U-Net variant with standard convolutions. Performance was assessed using quantitative metrics such as Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Fréchet Inception Distance (FID). The results showed that our model outperformed these benchmarks, achieving higher PSNR and SSIM scores, lower MAE and RMSE values, and a significantly reduced FID, indicating superior image quality and closer resemblance to ground-truth IHC images. Furthermore, the integration of depthwise separable convolutions led to a notable decrease in inference time and model size, improving its feasibility for clinical applications. These findings highlight the potential of our method as a significant advancement in virtual IHC staining, offering improved accuracy, efficiency, and suitability for broader clinical use.