IHCScoreGAN: An unsupervised generative adversarial network for end-to-end ki67 scoring for clinical breast cancer diagnosis

Carl Molnar, Thomas E. Tavolara, Christopher A. Garcia, David S. McClintock, Mark D. Zarella, Wenchao Han
Proceedings of The 7nd International Conference on Medical Imaging with Deep Learning, PMLR 250:1011-1025, 2024.

Abstract

Ki67 is a biomarker whose activity is routinely measured and scored by pathologists through immunohistochemistry (IHC) staining, which informs clinicians of patient prognosis and guides treatment. Currently, most clinical laboratories rely on a tedious, inconsistent manual scoring process to quantify the percentage of Ki67-positive cells. While many works have shown promise for Ki67 quantification using computational approaches, the current state-of-the-art methods have limited real-world feasibility: they either require large datasets of meticulous cell-level ground truth labels to train, or they provide pre-trained weights that may not generalize well to in-house data. To overcome these challenges, we propose IHCScoreGAN, the first unsupervised deep learning framework for end-to-end Ki67 scoring without the need for any ground truth labels. IHCScoreGAN only requires IHC image samples and unpaired synthetic data, yet it learns to generate colored cell segmentation masks while simultaneously predicting cell center point and biomarker expressions for Ki67 scoring, made possible through our novel dual-branch generator structure. We validated our framework on a large cohort of 2,136 clinically signed-out cases, yielding an accuracy of 0.97 and an F1-score of 0.95 and demonstrating substantially better performance than a pre-trained state-of-the-art supervised model. By removing ground truth requirements, our unsupervised technique constitutes an important step towards easily-trained Ki67 scoring solutions which can train on out-of-domain data in an unsupervised manner.

Cite this Paper

BibTeX
@InProceedings{pmlr-v250-molnar24a, title = {IHCScoreGAN: An unsupervised generative adversarial network for end-to-end ki67 scoring for clinical breast cancer diagnosis}, author = {Molnar, Carl and Tavolara, Thomas E. and Garcia, Christopher A. and McClintock, David S. and Zarella, Mark D. and Han, Wenchao}, booktitle = {Proceedings of The 7nd International Conference on Medical Imaging with Deep Learning}, pages = {1011--1025}, year = {2024}, editor = {Burgos, Ninon and Petitjean, Caroline and Vakalopoulou, Maria and Christodoulidis, Stergios and Coupe, Pierrick and Delingette, Hervé and Lartizien, Carole and Mateus, Diana}, volume = {250}, series = {Proceedings of Machine Learning Research}, month = {03--05 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v250/main/assets/molnar24a/molnar24a.pdf}, url = {https://proceedings.mlr.press/v250/molnar24a.html}, abstract = {Ki67 is a biomarker whose activity is routinely measured and scored by pathologists through immunohistochemistry (IHC) staining, which informs clinicians of patient prognosis and guides treatment. Currently, most clinical laboratories rely on a tedious, inconsistent manual scoring process to quantify the percentage of Ki67-positive cells. While many works have shown promise for Ki67 quantification using computational approaches, the current state-of-the-art methods have limited real-world feasibility: they either require large datasets of meticulous cell-level ground truth labels to train, or they provide pre-trained weights that may not generalize well to in-house data. To overcome these challenges, we propose IHCScoreGAN, the first unsupervised deep learning framework for end-to-end Ki67 scoring without the need for any ground truth labels. IHCScoreGAN only requires IHC image samples and unpaired synthetic data, yet it learns to generate colored cell segmentation masks while simultaneously predicting cell center point and biomarker expressions for Ki67 scoring, made possible through our novel dual-branch generator structure. We validated our framework on a large cohort of 2,136 clinically signed-out cases, yielding an accuracy of 0.97 and an F1-score of 0.95 and demonstrating substantially better performance than a pre-trained state-of-the-art supervised model. By removing ground truth requirements, our unsupervised technique constitutes an important step towards easily-trained Ki67 scoring solutions which can train on out-of-domain data in an unsupervised manner.} }
Endnote
%0 Conference Paper %T IHCScoreGAN: An unsupervised generative adversarial network for end-to-end ki67 scoring for clinical breast cancer diagnosis %A Carl Molnar %A Thomas E. Tavolara %A Christopher A. Garcia %A David S. McClintock %A Mark D. Zarella %A Wenchao Han %B Proceedings of The 7nd International Conference on Medical Imaging with Deep Learning %C Proceedings of Machine Learning Research %D 2024 %E Ninon Burgos %E Caroline Petitjean %E Maria Vakalopoulou %E Stergios Christodoulidis %E Pierrick Coupe %E Hervé Delingette %E Carole Lartizien %E Diana Mateus %F pmlr-v250-molnar24a %I PMLR %P 1011--1025 %U https://proceedings.mlr.press/v250/molnar24a.html %V 250 %X Ki67 is a biomarker whose activity is routinely measured and scored by pathologists through immunohistochemistry (IHC) staining, which informs clinicians of patient prognosis and guides treatment. Currently, most clinical laboratories rely on a tedious, inconsistent manual scoring process to quantify the percentage of Ki67-positive cells. While many works have shown promise for Ki67 quantification using computational approaches, the current state-of-the-art methods have limited real-world feasibility: they either require large datasets of meticulous cell-level ground truth labels to train, or they provide pre-trained weights that may not generalize well to in-house data. To overcome these challenges, we propose IHCScoreGAN, the first unsupervised deep learning framework for end-to-end Ki67 scoring without the need for any ground truth labels. IHCScoreGAN only requires IHC image samples and unpaired synthetic data, yet it learns to generate colored cell segmentation masks while simultaneously predicting cell center point and biomarker expressions for Ki67 scoring, made possible through our novel dual-branch generator structure. We validated our framework on a large cohort of 2,136 clinically signed-out cases, yielding an accuracy of 0.97 and an F1-score of 0.95 and demonstrating substantially better performance than a pre-trained state-of-the-art supervised model. By removing ground truth requirements, our unsupervised technique constitutes an important step towards easily-trained Ki67 scoring solutions which can train on out-of-domain data in an unsupervised manner.
APA
Molnar, C., Tavolara, T.E., Garcia, C.A., McClintock, D.S., Zarella, M.D. & Han, W.. (2024). IHCScoreGAN: An unsupervised generative adversarial network for end-to-end ki67 scoring for clinical breast cancer diagnosis. Proceedings of The 7nd International Conference on Medical Imaging with Deep Learning, in Proceedings of Machine Learning Research 250:1011-1025 Available from https://proceedings.mlr.press/v250/molnar24a.html.

Related Material