Deep learning in cancer pathology: a new generation of clinical biomarkers.
Echle A, Rindtorff NT, Brinker TJ, Luedde T, Pearson AT, Kather JN. · 2021
Read Deep learning in cancer pathology: a new generation of clinical biomarkers. by Echle A, Rindtorff NT, Brinker TJ, Luedde T, Pearson AT, Kather JN. (2021) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc7884739.txt.
Introduction
Echle and co-authors present a review of deep learning in cancer pathology. Cancer care leans on molecular tests to predict how a tumour will behave and which drug may work. Those tests cost time and money and need extra tissue. The review asks what image-based learning can add.
Routine tumour slides already hold rich information that standard review does not use. Deep learning systems can sort, grade, and subtype tumours from those images. More advanced systems go further. They infer molecular features, predict survival, and even forecast response to a therapy. The authors stress that such tools need strict testing in real clinics before they guide decisions. The review gives a clear tour of a field that could reshape how biomarkers are found.
Catalog record
| CREATOR | Echle A, Rindtorff NT, Brinker TJ, Luedde T, Pearson AT, Kather JN. |
|---|---|
| YEAR | 2021 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 59 KB |
| LICENSE | CC BY 4.0 |
| RIGHTS | CC BY |
| SOURCE | Europe PMC |
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Deep learning in cancer pathology: a new generation of clinical biomarkers By Amelie Echle, Niklas Timon Rindtorff, Titus Josef Brinker, Tom Luedde, Alexander Thomas Pearson, Jakob Nikolas Kather Published in: British Journal of Cancer DOI: 10.1038/s41416-020-01122-x ABSTRACT Abstract Clinical workflows in oncology rely on predictive and prognostic molecular biomarkers. However, the growing number of these complex biomarkers tends to increase the cost and time for decision-making in routine daily oncology practice; furthermore, biomarkers often require tumour tissue on top of routine diagnostic material. Nevertheless, routinely available tumour tissue contains an abundance of clinically relevant information that is currently not fully exploited. Advances in deep learning (DL), an artificial intelligence (AI) technology, have enabled the extraction of previously hidden information directly from routine histology images of cancer, providing potentially clinically useful information. Here, we outline emerging concepts of how DL can extract biomarkers directly from histology images and summarise studies of basic and advanced image analysis for cancer histology. Basic image analysis tasks include detection, grading and subtyping of tumour tissue in histology images; they are aimed at automating pathology workflows and consequently do not immediately translate into clinical decisions. Exceeding such basic approaches, DL has also been used for advanced image analysis tasks, which have the potential of directly affecting clinical decision-making processes. These advanced approaches include inference of molecular features, prediction of survival and end-to-end prediction of therapy response. Predictions made by such DL systems could simplify and enrich clinical decision-making, but require rigorous external validation in clinical settings. Subject terms: Targeted therapies, Tumour biomarkers, Computational science, Cancer imaging BACKGROUND Decision-making processes in oncology today no longer rely on workflows that are linear and straightforward; rather, with the availability of an ever-increasing number of biomarkers, these flowcharts resemble intricate trees with numerous branches, which consequently increase the complexity of treatment recommendations for solid tumours. Currently used molecular biomarkers in these oncology workflows can be prognostic or predictive. Prognostic biomarkers allow the categorisation of patients according to their risk of disease progression or death and, accordingly, can be used to adjust treatment intensity for individual patients. For example, in stage II colorectal cancer (CRC), microsatellite instability (MSI) is a prognostic biomarker; if MSI is detected, a lower treatment intensity of adjuvant chemotherapy can be used due to the inherently better prognosis of these patients.1 By contrast, predictive biomarkers enable a particular targeted treatment to be chosen for a specific patient group. For example, in treatment-refractory stage IV CRC, MSI is an FDA-approved biomarker for immune-checkpoint-inhibitor-based immunotherapy.2 In this case, the detection of MSI correlates with the likelihood of a positive therapeutic response, making MSI a strong predictive biomarker in this setting. Similarly, in breast cancer, the detection of HER2 positivity3 makes patients eligible for treatment with anti-HER2 agents, thus acting as a strong predictive biomarker in this disease.4 The choice of treatment for non-small-cell lung cancer (NSCLC) is influenced by a high number of molecular biomarkers,5 with oncogenic mutations in the gene encoding epidermal growth factor receptor (EGFR) and other genes, gene fusions of anaplastic lymphoma kinase (ALK) or other drivers and the overexpression of programmed cell death ligand 1 (PD-L1)6 being part of the standard-of-care molecular panel required for routine treatment of advanced or metastatic disease.7 It is clear, then, that the rapidly increasing number and clinical importance of molecular biomarkers in routine clinical practice allows cancer treatments to be tailored more specifically according to the genetic make-up of a particular tumour; consequently, however, the cost, turnaround time and tissue requirements in routine workflows also increase.8,9 The design of clinical trials for new therapeutic agents in solid tumours is increasingly coupled to predictive biomarkers. In addition to highly prevalent molecular features, many Phase 2 and 3 trials carried out over the past ~5 years have focused on rare molecular subpopulations of solid tumours, such as those with MSI,10 homologous repair deficiency11 and fusion-driven tumours across cancer types.12–14 As mentioned above, MSI is used as a predictive biomarker for immunotherapy, while homologous repair deficiency tumours are effectively targetable by inhibitors of poly ADP-ribose polymerase (PARP), and fusion-driven tumours respond exceptionally well to molecularly targeted therapy. However, with the prevalence of these genotypes ranging between 1% and 10% in real-world populations, screening potential participants for these trials is costly and hampered by the limited availability of molecular assays. So, despite the increased number of prognostic and predictive biomarkers enabling a more nuanced treatment of cancer patients, the complexity of clinical decision-making processes increasingly becomes an issue in clinical routine and clinical trial recruitment. Although most new biomarkers in oncology are based on molecular biology assays, advances in deep learning (DL) are facilitating the extraction of otherwise hidden information directly from routinely available data. DL is a method in the realm of artificial intelligence (AI) that makes use of artificial neural networks to identify recurring patterns in complex datasets. Image data in particular has a high information density, making it ideal for analysis with DL techniques. Indeed, DL-based image analysis has broad applications in multiple fields of modern medicine that involve image data: in radiology, DL performs repetitive tasks with human-like, or super-human, performance, such as tumour detection or organ segmentation on computer tomography (CT) images. To date, more than a dozen DL methods are approved for clinical use in radiology by the FDA—for example, DL-based analysis of CT data was carried out in a 2019 lung cancer screening trial,15 and evidence on the clinical usefulness of these methods is quickly mounting. Magnetic resonance imaging (MRI) data, which contain much more information than CT data, are also amenable for DL-based mining,16 and DL has also shown robust results for non-radiology tasks such as the analysis of real-time endoscopy images17,18 and skin cancer detection in dermoscopy images.19,20 Compared to these imaging modalities, however, histology is a ubiquitous image source with a remarkable information density that can be derived from routine clinical practice. Being much larger than radiological images in terms of pixels, images from histology slides carry much more information: millions of different cells can be seen in a histology slide and their morphology and spatial arrangement carry much more information than other medical images. Even the size of a whole chest CT dataset does not get close to the size of the dataset from one histological whole- slide image derived from the tumour of the same patient when measured in pixels (Fig. 1a, b). This high information density makes histological images an attractive source for DL-based biomarker extraction. [Figure: Consensus pipeline of deep learning in pathology.a Routine histology image of lung cancer (from The Cancer Genome Atlas (TCGA) and The Cancer Imaging Archive (TCIA)). b Size comparison (in terms of pixels) of a chest CT scan of the same patient. c Consensus image-processing pipeline. First, either the whole slide or just the tumour region is tessellated into smaller image tiles. d These tiles comprise an image library, similar to the library preparation (prep.) in genome sequencing. e Tiles are preprocessed to achieve rotational constancy and augment the dataset. f Deep- learning classifiers are developed and deployed by splitting the patient cohort into a training and testing set, by using cross-validation or by having multiple cohorts available for training and testing. g Ideally,