Prediction of Ki67 scores from H&E stained breast cancer sections using convolutional neural networks
Keywords: Ki67, breast cancer, convolutional neural networks, Cycle-GAN, registration, histopathology, digital pathology
TL;DR: We compared four CNN based modelling approaches to directly predict Ki67 scores from WSIs of H&E stained histology sections in a dataset consisting of 126 breast cancer cases.
Abstract: Ki67 is an established marker of proliferation in breast cancer, but currently has limited clinical value due to limitations of the analytical validity of immunohistochemistry (IHC) -based Ki67 scoring. While the inter-assessor variability of scoring can be improved through image analysis software, Ki67 IHC also suffers from a lack of standardized staining protocols and is not part of routine pathology workflow in most countries. This could potentially be alleviated through directly predicting Ki67 scores from routine hematoxylin and eosin (H\&E) stained whole-slide-images (WSIs). We compared four different deep learning based approaches to predict Ki67 scores from routine H\&E stained WSIs in a dataset that consists of matched H\&E and Ki67 WSIs from 126 breast cancer patients, resulting in a Spearman correlation between WSI cancer ROI averages of 0.546 for the best performing model in a 5-fold cross-validation (CV). These findings suggest that it is possible to predict the Ki67 score from H\&E stained WSIs, but validation in a larger cohort is required to meaningfully distinguish the performance of the methods that were investigated.
Paper Type: validation/application paper
Primary Subject Area: Application: Histopathology
Secondary Subject Area: Integration of Imaging and Clinical Data
Paper Status: original work, not submitted yet
Source Code Url: The prediction modelling code relies on internal tooling and is therefore not going to be made available publicly at this stage. Cycle-GAN code is available here: https://junyanz.github.io/CycleGAN/
Data Set Url: The data set cannot be made publicly available at this time due to confidentiality and personal data protection requirements.
Registration: I acknowledge that publication of this at MIDL and in the proceedings requires at least one of the authors to register and present the work during the conference.
Authorship: I confirm that I am the author of this work and that it has not been submitted to another publication before.