Towards computationally efficient prediction of molecular signatures from routine histology images

Abstract

Identification of actionable genomic alterations in diagnostic tissue samples provides key information for personalised cancer treatment. However, current diagnostic tests used to predict the status of molecular pathways from standard histopathology material are commonly tissue destructive, generate relevant costs, and can take hours or even days to return conclusive results. Relating molecular alterations to digital image information derived from standard haematoxylin and eosin-stained histopathology slides has become a task that machine learning models are able to solve across multiple cancer types (appendix).1Coudray N Ocampo PS Sakellaropoulos T et al.Classification and mutation prediction from non-small cell lung cancer histopathology images using deep learning.Nat Med. 2018; 24: 1559-1567Crossref PubMed Scopus (922) Google Scholar, 2Echle A Grabsch HI Quirke P et al.Clinical-grade detection of microsatellite instability in colorectal tumors by deep learning.Gastroenterology. 2020; 159: 1406-1416.e11Summary Full Text Full Text PDF PubMed Scopus (83) Google Scholar, 3Kather JN Heij LR Grabsch HI et al.Pan-cancer image-based detection of clinically actionable genetic alterations.Nat Can. 2020; 1: 789-799Crossref PubMed Scopus (127) Google Scholar, 4Popovici V Budinská E Dušek L Kozubek M Bosman F Image-based surrogate biomarkers for molecular subtypes of colorectal cancer.Bioinformatics. 2017; 33: 2002-2009Crossref PubMed Scopus (10) Google Scholar, 5Sirinukunwattana K Domingo E Richman SD et al.Image-based consensus molecular subtype (imCMS) classification of colorectal cancer using deep learning.Gut. 2021; 70: 544-554Crossref PubMed Scopus (46) Google Scholar, 6Veta M Heng YJ Stathonikos N et al.Predicting breast tumor proliferation from whole-slide images: the TUPAC16 challenge.Med Image Anal. 2019; 54: 111-121Crossref PubMed Scopus (98) Google Scholar, 7Weitz P Wang Y Kartasalo K et al.Transcriptome-wide prediction of prostate cancer gene expression from histopathology images using co-expression based convolutional neural networks.arXiv. 2021; (https://arxiv.org/abs/2104.09310 (preprint) published online April 19)Google Scholar For digitised pathology labs, this technology promises the introduction of support tools that will be able to flag the presence of molecular alterations from histology slides in the intermediate future. Such image-based molecular classification algorithms have the potential to facilitate the cost-effective preselection of patients for molecular testing and to screen for hereditary cancer syndromes. In future development stages, image-based classification tools trained on outcome data could also directly aid clinical stratification and help to achieve optimal treatment decisions in difficult clinical scenarios.5Sirinukunwattana K Domingo E Richman SD et al.Image-based consensus molecular subtype (imCMS) classification of colorectal cancer using deep learning.Gut. 2021; 70: 544-554Crossref PubMed Scopus (46) Google Scholar Slide-level representation learning is an important area of research in computational pathology because discriminative morphological features need to be automatically discovered and located to make the desired prediction. However, due to their large size, processing whole-slide images comes with technical challenges that so far limit the efficacy of conventional computational approaches and impair the progress towards clinical deployment. As existing computational resources do not enable the direct processing of whole-slide images with standard deep learning architectures, current designs typically entail a tradeoff between the quality of the learned representation and how much computational complexity to use. To answer this technical challenge, the original task of classifying gigapixel images can be simplified into a weakly supervised learning problem by splitting whole-slide images into tiles, each of which is then processed individually. This approach provides class predictio