According to the WHO’s records, the incidence of cancer worldwide in the year 2020 was 19 million. This figure is expected to rise in the coming years. An important step in preventing cancer progression and enabling improved patient outcomes is early screening. The current gold standard for cancer screening requires pathologists to examine slides of stained tissue biopsies from the suspected tumor-affected regions and identify key morphological features characteristic of tumors. Apart from being a tedious, time-consuming activity, this form of screening also requires a high level of human skill to be able to spot minute but important differences in tissue architecture. Tumors also have to be graded by severity and classified by subtypes in many cases.

Deep Learning to Lend a Hand

Deep learning is a branch of machine learning and makes use of artificial neural networks to pick out recurring patterns from complex datasets. Scientists have taken advantage of this ability to simplify processes in medicine and healthcare, notably in the case of cancer screening and prognosis.

Digital Pathology

Since stained tissue biopsy images contain millions of cells and provide a wealth of data, deep learning algorithms have been trained to become digital pathologists. This technology can now be used to assess whether a patient has cancer or not, following image analysis of the biopsies, with a high level of accuracy. Additionally, tumor grading and subtyping have also been automated similarly. These basic applications have gained FDA approval, greatly simplifying a pathologists’ workflow. While digital pathology has reduced human effort and turnaround time, it’s still a pretty basic application of deep learning. Its potential can be further harnessed for advanced applications such as genotype and survival prediction.

Cancer Genotyping

Mutations that drive cancer, called oncogenic driver mutations, bring about morphological cellular changes and can affect aspects such as the texture of the cytoplasm and the nucleus, size and shape, etc. They also exert their effects on neighbouring cells and induce secondary changes in tissue architecture. A novel and exciting application of deep learning is that it can be taught to identify the driver mutation just by image analysis, as described by Coudray et al. Their paper showed that it was possible to automate cancer genotyping based on histology in the case of mutations in TP53, EGFR, and STK11, among others.

Survival Prediction

Treatment modalities and survival predictions are based on a host of criteria including patient age, gender, stage of cancer, underlying conditions, and histology risk factors. Deep learning technology can be taught to predict patient survival by combining image analysis with clinical parameters. A paper by Bychkov et al. predicted the 5-year survival of patients affected with colorectal cancer based only on histology images.

Looking Forward

These advanced applications, however, have only been explored in proof-of-concept studies. A lot of work remains to be done before they can be employed by oncologists in real-world scenarios. Automated survival prediction and genotyping still lag in accuracy as compared to wet-lab techniques that are currently in use. Researchers are focused on improving accuracy by training their algorithms on larger datasets to receive clinical validation and FDA approval.

Nevertheless, deep learning exhibits tremendous potential in simplifying and automating various aspects of pathology and oncology.

References:

1. https://www.nature.com/articles/s41416-020-01122-x
2. https://www.nature.com/articles/s41571-019-0252-y 
3. https://www.nature.com/articles/s41591-018-0177-5
4. https://www.nature.com/articles/s41598-018-21758-3

<!-- SharpSpring Form for Newsletter Subscription_Website blogs  -->
<script type="text/javascript">
    var ss_form = {'account': 'MzawMDE3tjQxBAA', 'formID': 'M0wxMzQ2sTDXNUtOMtA1STQy1bVMSknWTU0yTjMwMjI2S0wzAAA'};
    ss_form.width = '100%';
    ss_form.height = '1000';
    ss_form.domain = 'app-3QNHZALW1I.marketingautomation.services';
    // ss_form.hidden = {'field_id': 'value'}; // Modify this for sending hidden variables, or overriding values
    // ss_form.target_id = 'target'; // Optional parameter: forms will be placed inside the element with the specified id
    // ss_form.polling = true; // Optional parameter: set to true ONLY if your page loads dynamically and the id needs to be polled continually.
</script>
<script type="text/javascript" src="https://koi-3QNHZALW1I.marketingautomation.services/client/form.js?ver=2.0.1"></script>