Efficient deep neural networks for cancer detection on histopathology combining attention and image downsampling

TL;DR

This paper presents an efficient deep learning model for detecting colorectal cancer in histopathology images, using attention mechanisms and reduced image resolution to maintain accuracy while lowering computational costs.

Contribution

The novel approach combines attention and image downsampling to achieve high accuracy in cancer detection with reduced computational resources.

Findings

The model achieved an F1-Score of 0.96, a Matthews correlation coefficient of 0.92, and an AUC of 0.99 on test data.

Working at 4 μm/pix resolution maintained high performance while reducing computational costs.

The model was validated on over 1300 patients and tested on The Cancer Genome Atlas dataset.

Abstract

Pathology diagnosis of colorectal cancer is time-consuming and requires a high level of expertise. However, it is an essential step towards establishing the adequate treatment. The need to analyse a large number of these histopathological images calls for automatic tools capable of aiding pathologists in this arduous task. Deep learning techniques, together with the wealth of data available nowadays, provide a promising candidate for such job. Adopting state-of-the-art artificial intelligence algorithms, we developed a model to accurately detect colorectal cancer in digitalised histopathological whole-slide images. Our end-to-end approach uses the principles of multiple-instance learning combined with deep convolutional neural networks in order to fully leverage the information contained within each image and make robust predictions at the patient’s level. The model also allows to highlight the areas in the slide most likely to harbour tumour tissue. Given the finite computational resources available, working at maximum resolution can be detrimental. Therefore, we explored the impact of lowering the working image resolution. The algorithms were trained and validated on a subset of more than 1300 patients of the Molecular Epidemiology of Colorectal Cancer study with histopathology images available. These images gave rise to \documentclass[12pt]{minimal}
\usepackage{amsmath}
\usepackage{wasysym}
\usepackage{amsfonts}
\usepackage{amssymb}
\usepackage{amsbsy}
\usepackage{mathrsfs}
\usepackage{upgreek}
\setlength{\oddsidemargin}{-69pt}
\begin{document}$$>10^5$$\end{document} tiles of \documentclass[12pt]{minimal}
\usepackage{amsmath}
\usepackage{wasysym}
\usepackage{amsfonts}
\usepackage{amssymb}
\usepackage{amsbsy}
\usepackage{mathrsfs}
\usepackage{upgreek}
\setlength{\oddsidemargin}{-69pt}
\begin{document}$$256\times 256$$\end{document} pixels each. Once we identified the best-performing model we put it to the test on images from The Cancer Genome Atlas. We obtained the best outcomes working at 4 μm/pix, achieving the following metrics on the test dataset: F1-Score of 0.96, a Matthews correlation coefficient of 0.92 and an area under the receiver operating characteristic curve of 0.99. These results are exceptional and prove that computational costs can be reduced while keeping the performance up to standard.

Figures