Tri-Model Integration: Advancing Breast Cancer Immunohistochemical Image Generation through Multi-Method Fusion
Immunohistochemical (IHC) staining is a crucial technique for diagnosing and formulating treatment plans for breast cancer, particularly by evaluating the expression of biomarkers like human epidermal growth factor receptor-2. However, the high cost and complexity of IHC staining procedures have driven research toward generating IHC-stained images directly from more readily available Hematoxylin and Eosin-stained images using image-to-image (I2I) translation methods. In this work, we propose a novel approach that combines the predictive capabilities of three state-of-the-art I2I models to enhance the quality and reliability of synthetic IHC images. Specifically, we designed a Convolutional Neural Network that takes as input a four-dimensional input comprising the outputs of three distinct models (each contributing an IHC prediction, which is an RGB three-dimensional output for each) and produces a final consensus image through a fusion mechanism. This ensemble method leverages the strengths of each individual model, leading to more robust and accurate IHC image generation. Extensive experiments on the BCI dataset demonstrate that our approach outperforms existing single-model methods, achieving superior Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) metrics.
| Model | PSNR (dB) | SSIM |
|---|---|---|
| CycleGAN (Unsupervised) [12] | 16.203 | 0.373 |
| Pix2PixHD [9] | 19.634 | 0.471 |
| Pyramid Pix2Pix [2] | 21.160 | 0.477 |
| BCIStainer [4] | 22.929 | 0.559 |
| CUT with PatchNCE [10] | 17.927 | 0.555 |
| Proposed Method | 23.206 | 0.614 |
The BCI dataset can be accessed through bci page
The three models we used in our architecture can be found here:
BCIStainer: git-hub link
Pyramid Pix2Pix: git-hub link
(CUT) with Supervised PatchNCE Loss: git-hub link
First download the BCI dataset and use the instructions for each of the three models to evaluate the IHC images for the train and test sections of the BCI dataset for each of the models.
after evaluating the IHC databases for each method, follow the markdowns on the BCI_fusion.ipynb file to use our fusion method for training or evaulation.
pip install -r requirements.txtfor maximum convenience we implemented all of the code in a single jupyter notebook file, which can be used on any platform (windows, linux, colab, ...). The markdowns in the file guide you thoroughly on how to train or evaluate the model.
The trained weights are available on google drive which you can use to avoid training from scratch.
The code is run on Nvidia GTX 1080.
[2] S. Liu, C. Zhu, F. Xu, X. Jia, Z. Shi, and M. Jin. Bci: Breast cancer immunohistochemical image generation through pyramid pix2pix, 2022.
[4] C. Zhu, S. Liu, Z. Yu, F. Xu, A. Aggarwal, G. Corredor, A. Madabhushi, Q. Qu, H. Fan, F. Li, Y. Li, X. Guan, Y. Zhang, V. K. Singh, F. Akram, M. M. K. Sarker, Z. Shi, and M. Jin. Breast cancer immunohistochemical image generation: a benchmark dataset and challenge review, 2023.
[9] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8798–8807, 2018.
[10] Lingke Kong, Chenyu Lian, Detian Huang, Zhenjiang Li, Yanle Hu, and Qichao Zhou. Breaking the dilemma of medical image-to-image translation. arXiv preprint arXiv:2110.06465, 2021.
[12] J.-Y. Zhu, T. Park, P. Isola, and A. A. Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 2242–2251, 2017.