Deep Learning plays a vital role in the early detection of cancer. A study published by NVIDIA showed that deep learning drops error rate for breast cancer diagnoses by 85%. This was the inspiration for Co-Founders Jeet Raut and Peter Njenga when they created AI imaging medical platform Behold.ai. Raut’s mother was told that she no longer had breast cancer, a diagnosis that turned out to be false and that could have cost her life.
In addition to being the second leading cause of death (responsible for 8.8 million deaths worldwide in 2015), cancer also has significant and increasing impacts on economy. In 2010, the total annual cost of cancer was estimated at around $1.6 trillion. But the good news is that early detection can save not only billions of dollars but countless lives. In a TEDx 2014 talk, CEO and Founder of Enlitic Jeremy Howard said, “If you detect cancer early, your probability of survival is 10 times higher.”
Fortunately, deep learning has shown capabilities in achieving higher diagnostic accuracy results in comparison to many domain experts. While this may be an issue of contention with physicians, for many would-be victims the technology can’t come soon enough.
In this article, we’ve provided several examples of deep learning in oncology. Though this list is by no means complete, it gives an indication of the long-ranging impact of deep learning on the oncology industry today and in the near future.
Current Applications of Deep Learning in Oncology
Cancer Detection From Gene Expression Data
Gene expression data is very complex due to its high dimensionality and complexity, making it challenging to use such data for cancer detection. Researchers from Oregon State University were able to use deep learning for the extraction of meaningful features from gene expression data, which in turn enabled the classification of breast cancer cells. They have used the technology to extract genes considered useful for cancer prediction, as well as potentially useful cancer biomarkers, for the detection of breast cancer.
Researchers in China also developed DeepGene, an advanced cancer type classifier based on deep learning that addresses the obstacles in existing somatic point mutation based cancer classification (SMCC) studies. Results showed that DeepGene outperforms three widely adopted existing classifiers, as it was able to extract the high-level features between combinatorial somatic point mutations and specific cancer types.
Cancer Classification with Deep Neural Networks
A research paper published in Nature by Stanford University researchers has shown that their convolutional neural network (CNN) achieves performance on par with all tested experts when classifying skin cancer. Google’s CNN system has demonstrated the ability to identify deadline skin cancers at an accuracy rate on par with practitioners, potentially extending diagnosis reach outside the clinic and into service-based apps that have been popping up as mobile access expands worldwide. Yet another instance is a program developed at Case Western Reserve University that has been able to outperform physicians at brain cancer diagnoses.
Tumor Segmentation
Researchers from China have used deep learning for segmenting brain tumors in MR images, where it provided more stable results as compared to manually segmenting the brain tumors by physicians, which is prone to motion and vision errors.
A team led by Dr. Qi Zhang of Shanghai University found that deep learning can accurately differentiate between benign and malignant breast tumors on ultrasound shear-wave elastography (SWE), yielding more than 93% accuracy on the elastogram images of more than 200 patients.
Histopathologic Cancer Diagnosis
With the advent of personalized medicine, diagnostic protocols need to focus equally on efficiency and accuracy, thus increasing the workload and complexity of histopathologic (microscopic examination of tissue in order to study the manifestations of disease) in cancer diagnosis. This has led researchers in the Netherlands to use deep learning to improve the efficiency of histopathologic slide analysis, where the workload for pathologists is reduced and the objectivity of diagnoses is increased. The researchers concluded that deep learning could improve the efficacy of prostate cancer diagnosis and breast cancer staging.
In another case, Philips and LabPON are planning to create the world’s largest pathology database of annotated tissue images for deep learning. One of the things the database will provide is data for research and discovery to develop new insights in disease assessment, including cancer.
In an effort to accelerate cancer research, Oak Ridge National Laboratory (ORNL) researchers are applying deep learning toward automating information collected from cancer pathology reports that are documented across a nationwide network of cancer registry programs.
Tracking Tumor Development
Deep learning can be used to measure the size of tumors undergoing treatment and detect new metastases that might be overlooked. This is exactly what researchers from the Fraunhofer Institute for Medical Image Computing in Germany are doing. The more patient CT and MRI scans the deep learning algorithm reads, the more accurate it becomes, which is the core of deep learning technology.
Google’s deep learning tumor prediction heat maps
Google Research is also hard at work developing deep learning tools that can “naturally complement pathologists’ workflow.” They used images to train their deep learning algorithm Inception (aka GoogLeNet) to identify breast cancer tumors that have spread to adjacent lymph nodes. The algorithm reached a localization score of 89%, exceeding the 73% accuracy rate for pathologists.
Prognosis Detection
Prognosis provides an estimate of how serious or advanced the stage of cancer, and hence the chances of survival. Staging systems for cancer are critical for predicting the patient’s prognosis but suffer from limitations. Researchers in South Korea utilized deep learning to develop a prediction model for the prognosis of patients suffering from gastric cancer and undergoing treatment (i.e. gastrectomy). They found that deep learning showed superior survival predictive powers compared to other prediction models.
Deep learning for the quantification of tumor-infiltrating immune cells in breast cancer samples has also been used by researchers in Finland and Sweden. Immune cell infiltration in tumors is considered an emerging prognostic biomarker, and the gold standard for the quantification of immune cells in tissues is the visual assessment through a microscope. In this study, deep learning provided good discrimination of immune cell-rich areas.
Closing Thoughts on Deep Learning in Oncology
Many research-oriented entities are encouraging companies to innovate with machine and deep learning in the field of oncology, while others are publishing and making their research and insights on deep learning in oncology available to the public.
Kaggle is hosting a $1 million competition to improve lung cancer detection with machine learning. Elsevier is hosting a special issue on deep learning for computer aided cancer detection and diagnosis with medical imaging. Springer has published a book on machine learning in radiation oncology. The list goes on. Many startups — including Behold.ai, Enlitic, Insilico Medicine, Freenome, and other — are also joining in the global mission to stomp out cancer.
Early cancer detection and prognosis is one of several important healthcare areas where deep learning technology has been applied. The use of deep learning in oncology increases the chances that one day, machines may help researchers find a coveted cure and prevention methods for the development of cancer.
Image credit: Stanford News
