End-to-end deep neural network for ER, PR and Ki-67 stained WSI automatic proliferation index (PI) quantification for breast cancer tissue

Yahia Salhi, Joseph Rynkiewicz, Céline Bossard, Sanae Salhi, Jérôme Chetritt

2023

Abstract

Background & objectives

We develop an end-to-end pipeline based on a deep neural network for the quantification and assessment of nuclear biomarkers on stained WSI, i.e. ER, PR and Ki-67. The proliferation rate as well as hotspot detection are proposed.

Methods

A modified U-Net architecture with an adequate backbone network is investigated, trained and validated on 1656 tiles. Testing was carried on a left-out set of size 701 of the same database originating from 23 patients. The network allows the quantification of tumour cells, detection of hotspots and the quantification of the corresponding PI. The performance is investigated using external datasets.

Results

In this study, the model's performance in estimating ER, PR, and Ki-67 index scores was evaluated on a left-out dataset and two external datasets (n=23, 50, 130). The model outperformed existing frameworks on the same test data for the Ki-67 (n=23) with an accuracy of 95.96%. On the external datasets, i.e. AIDPATH (n=50) and IHP Group (n=130), the model achieved an averaged accuracy scores of 98.3% and 96.7%, respectively. Also, the hotspots for Ki-67 prediction were accurately inferred based on the prediction outcomes and those were compared to expert pathologists annotations. Additionally, the proposed end-to-end approach enables faster inference times compared to other available algorithms.

Conclusion

AI models for whole-slide image processing are usually task-specific and require complex processing pipelines. Unlike most algorithms that use a two-step approach to identify invasive areas and hot-spots before quantifying PI, the proposed end-to-end approach achieves this in a single step with promising results. This approach is easily adaptable to other biomarkers for prognosis and theranostics, as it doesn't require specialized intermediate models.

See DiaKwant^® Ki‑67

Try it on your cases ↗︎