About 12% of American women will develop breast cancer in their lifetimes. With a death rate of 16%, it is estimated that 42,690 US deaths in 2020 will be caused by breast cancer. That means that about 1 in 8 women will be diagnosed with breast cancer and about 1 in 6 of individuals diagnosed with breast cancer will die.
Despite significant R&D investments in finding a cure, cancer continues to confound specialists because it is an assorted disease caused by multiple factors that vary from case to case. To determine risk and treatment, doctors must consider a wealth of information from different modalities. Oncologists must synthesize factors ranging from the patient’s genome , the tumor’s genome, and text from their electronic health records (EHRs), to demographic information, images from pathology, and images from radiology.
It is a challenge for humans to synthesize this trove of data. In contrast, computers are well suited to cancer predictions that are based on complex data using artificial intelligence (AI). In this article, we’ll explore how bringing digital tools into oncology can make doctors more effective in the fight against cancer, with a specific focus on breast cancer.
Digital oncology in the prediction of breast cancer:
Coupling machine learning with the Gail model:
Data from EHRs and the genetic sequencing of patients can be used to predict the risk of developing cancer. For decades, the Breast Cancer Assessment Tool, or “Gail model,” has been used to predict a patient’s probability of developing breast cancer.
The Gail model gives a risk score based on EHR data such as hormone levels, age at first menstruation, race/ethnicity, and the number of first-degree relatives who developed cancer to score a patient’s likelihood of developing cancer. Machine learning (ML) algorithms can program themselves based on Gail model data to learn to make accurate predictions on breast cancer, eventually outperforming physicians who must manually assess these factors.
Predicting the probability of acquiring a mutation:
Digital oncology can also help uncover the underlying genetic risk for cancer that a patient may have. Mutation of BRCA1 and BRCA2, two tumor suppressor genes, are the most common culprits in hereditary breast and ovarian cancers. Mutations in a single letter of the genetic code, called single-nucleotide polymorphisms (SNPs), acquired elsewhere in the genome can also be risk factors in developing cancer. Using advanced genetic sequencing platforms, computers can analyze genetic sequencing data from a patient to measure the probability that they will acquire a mutation in a high-risk gene.
Put together, these data can allow us to split the population into high-risk and low-risk groups for breast cancer, thus allowing individuals in high-risks groups to take active steps to decrease their risk of developing cancer, such as undergoing earlier cancer screening and decreasing exposure to carcinogens.
More drastic measures include preventative mastectomies, which can reduce the risk of developing cancer by up to 95% in people with BRCA1 or BRCA2 mutations. Integrating data from multiple sources will allow doctors to make better decisions of what preventative measure to recommend based on a patient’s personal risk profile.
Still, a nascent field:
The use of AI in the prediction of breast cancer risk is an area of active research, although its use in the clinic is currently hindered by many technical factors. In addition, this technology needs to be extensively validated for accuracy. Nonetheless, the significant efforts being spent to push forward this approach mean that we can expect widespread AI-based breast cancer risk prediction in the clinic within the coming decade.
Digital oncology in diagnosis and prognosis of breast cancer:
Prevention is a crucial part of decreasing the mortality of breast cancer, and one of the key tools in our arsenal is early screening. Unfortunately, Dutch researchers have estimated that delayed cancer screening due to COVID-19 may result in an increase of up to 2.35 per 100,000 breast cancer–related deaths. Although telemedicine has risen to prominence during the pandemic and presents opportunities for remote delivery of care to cancer patients, most present uses are suited to consultations and follow-up visits, not screening mammograms.
Screening:
Earlier this year, mobile screening units were successfully piloted in Italy to permit early cancer screenings for individuals who could not use traditional services.
Clinical trials are also in place for noninvasive clinical procedures such as infrared imaging and ultrasound to enhance the available screening options. These technologies can be expected, with limited use, in the clinic within 5 years. Both should be supervised and used by expert pathologists who can verify the assessments.
Staging:
Once cancer is detected, tumor data can be integrated into data used for predictions to obtain a better projection of the patient’s disease evolution or prognosis. To stage a tumor at present, pathologists typically repeat the tedious process of imaging tissue mounted onto glass slides one by one with different specifications, and then count individual features to establish a diagnosis.
The examination of biopsies by pathologists will also be digitized. Existing digital slide-scanning platforms such as NanoZoomer cut image acquisition and analysis time to a fraction of the current amount. These automated platforms image all samples on glass slides at different depths and zoom levels and place them in a single file for fast reference. Once acquired, image processing ML algorithms such as segmentation can identify cells and their nuclei from images. This would greatly outperform manual image acquisition and cell counting, allowing pathologists to much more efficiently interpret results.
Digital oncology in the treatment of breast cancer:
Use of omics in predicting drug therapy responses:
The availability of genomic and proteomic data, combined with drug and treatment response predictions, can be used to personalize cancer treatment. Publicly available databases such as Genomics of Drug Sensitivity in Cancer and the Cancer Genome Atlas can be used to combine tumor mutation data with the tumor’s response to different drug treatments. ML algorithms can couple drug sensitivity data to the DNA sequences of the patient’s tumor genome to generate a personalized course of treatment. In effect, this process would find the most effective treatment for the patient’s particular cancer.
In addition to genomic data, the protein readout of a cell can also be used to predict drug response. Proteins are the workhorse molecules of the cell and the effectors of cell modifications arising from genetic mutations. Therefore, the proteomic profile is a more direct readout of cell function. Once proteomic databases are more widely available, they can also be incorporated into algorithms. For similar reasons, investigators have looked at metabolic and transcriptomic data as additional predictors of drug responses.
Biochip for cell-type analysis:
Researchers at the University of California, Irvine, have developed a new lab-on-a-chip that combines ML with microfluidics to determine cancer cell type. Oncologists can use cell-type classification to predict the response of patients to different treatments.
Dual-organ-on-a-chip system for detection of cancer drug-related heart damage:
Radiotherapy and chemotherapy can lead to heart damage. To track patients’ hearts in a noninvasive way, researchers developed a heart-breast dual-organ-on-a-chip system. The team at Terasaki Institute of Biomedical Innovation created a device with two connected chambers. One chamber contained cultured breast tumor tissue and the other contained cardiac tissue from the patient. The chambers connect by microfluidics to resemble the interaction of the tissues in the body. The researchers tracked the cultured cells for any changes in biomarkers. This allows for early detection of cancer treatment–related heart damage.
Coupling imaging with AI:
Imaging methods can also be used to predict breast cancer response to treatment. Researchers at Memorial Sloan Kettering have developed a new AI method that uses imaging data to predict a breast tumor’s response to chemotherapy. Using features extracted from radiology and MRI images, they could predict mutation status at oncogene HER2 and thus how well the patient would respond to anti-HER2 therapy.
New ML methods that allow oncologists to prescribe the exact radiation dose to a tumor based on data from computerized tomography (CT) scans are already being used in the clinic. This approach helps find the minimum effective dose of radiation and diminish some of the negative side effects of radiotherapy.
Limitations and considerations:
The implementation of ML and other AI-based approaches in oncology has been made possible by the ability to handle the computational demand for storage, handling, and analysis of very large and complex datasets. AI is already employed in imaging methods such as radiology and endoscopy, where computer-aided diagnostic and detection systems are used to segment medical images and detect key features.
Although digital tools have high potential in oncology, their adaptation to the clinic must be followed by exhaustive validation, since a mistake can be critical. As digital oncology is incorporated into clinical workflows, it will not replace the functions of pathologists, oncologists, and radiologists. These digital tools should rather serve to enhance and facilitate the work of these experts. It is critical that in the use of digital tools, doctors are trained to understand the ML algorithms, including the assumptions made and the variables at play, and not just use them as “black boxes.”
An important ethical barrier is the protection of private health data, since these algorithms need to be trained on large datasets of patient records in order to function optimally. Keeping patient data anonymous and secure is the responsibility of healthcare providers that adapt these methods in their practice and must be considered an integral part in the design of any digital tool.
Conclusions:
Digital tools to aid a doctor’s expertise are key to integrating complex information to create a personalized set of recommendations for each patient. This will directly lead to earlier detection from risk factors based on EHR and genetic sequencing data, faster and more precise diagnosis with digital pathology, and more targeted treatment with personalized treatments and tailored radiation doses.
AI in oncological image analysis already has limited use in the clinic. Other image analysis and pathology tools mentioned in this article may be deployed in the next 5 years, since they have a precedent, whereas more complex techniques such as the use of EHR and genomic data may be seen in the clinic within 10 years. Although these tools need to be exhaustively tested before widespread adoption in the clinic, once these barriers are overcome, digital oncology will become the new normal for cancer care.
