End-to-End Deep Neural Network for Ki-67 Stained WSI Automatic Hotspot Detection and Proliferation Index (PI) Quantification for Breast Cancer Tissue

01 MAR 2023

Abstract

Background

The advent of digitalisation of immunostained slides and recent progress in computer vision based on artificial intelligence emphasised the development of computer‐assisted quantification, assessment and analysis of whole-slide images (WSI). In this paper, we shed light on the nuclear protein Ki‐67 assessment, which is one of the main prognostic factors for tumor progression and treatment in breast cancer. The main objective is to propose an end-to-end pipeline based on a deep neural network for the hotspot detection on Ki-67 stained WSI as well as proliferation rate quantification.

Design

We developed a supervised deep neural network on histopathology annotated squared images (tiles) of size 256 pixels of the publicly available dataset SHIDC-BC-Ki-67, which consists of 2356 images labelled by expert pathologists. For each tile, cell annotations were available in three different classes: immunopositive, immunonegative, and tumor infiltrating lymphocytes (TIL). A 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 cells, detection of hotspots and the quantification of the corresponding PI. Finally, the performance of the model is investigated and compared to the original PathoNet architecture, among others.

Results

Ki-67 index score estimation was evaluated on the basis of root mean squared errors (RMSE) at the tile and and patient level. The proposed architecture achieved an RMSE of 0.035 at the patient level, which outperforms the benchmark, on the same data split, which has an RMSE of 0.048. The model also outperforms other state of the art frameworks. The hotspot is implicitly and accurately inferred on the basis of outcomes of this prediction. Regarding the classification of the patients with respect to the cut-off ranges: between 0 and 16%, 17% and 30% and higher than 31%, our approach achieves an accuracy of 95.96% over 23 patients on the left-out SHIDC-BC-Ki-67 database. The pipeline is also challenged using an external set of 65 cases of breast cancer, which were independently screened by experienced pathologists.

Conclusion

Development of artificial intelligence models for whole-slide images treatment is generally task-specific and is based on casting possibly complex processing pipelines. Unlike most of the available algorithms which are based on a two-step approach that first identifies the hotspot and then quantifies the PI, the proposed end-to-end approach achieves this on a single end-to-end step with promising performance. This model can be easily generalized to adapt to other prognostic and/or theranostic nuclear biomarkers in a quick and straightforward manner as it does not require stacking of specialized intermediate models.

Read the full paper

See DiaKwant^® Ki‑67

Try it on your cases ↗︎