Cancer is a leading cause of death and affects millions of lives every year. Its early detection could help to increase the survival of many lives1 in addition to saving billions of dollars.2 Most of the healthcare data are obtained from ‘omics’ (such as genomics, transcriptomics, proteomics, or metabolomics), clinical trials, research and pharmacological studies. Such data are highly complex, variable, and multidimensional. Sometimes such data are available from incompatible data sources. Unfortunately, the bulk of this data remains underutilized and could be used for biomarker identification and drug discovery.
Deep learning (DL) is a member of the larger machine learning (ML) and artificial intelligence (AI) family. It has been applied in many fields like computer vision, speech recognition, natural language processing, object detection, and audio recognition.3 Deep learning architectures, including deep neural networks (DNNs) and recurrent neural networks (RNNs), have been persistently improving the state of the art in drug discovery and disease diagnosis.4 Deep learning has the potential to achieve good accuracy for the diagnosis of various types of cancers, such as breast, colon, cervical, and lung cancer. It builds an efficient algorithm based on multiple processing layers of neurons5 (see Figure 1). However, the output (i.e. accuracy) of any deep learning model depends on multiple factors including, but not limited to, data type (numeric, text, image, sound, video), data size, architecture, and data ETL (extract, transform, load) and so on.
In this article we explore how deep learning has been successfully applied to potential areas of oncology (the study of cancer diagnosis and treatment). It provides an insight into deep learning for medical and paramedical professionals, educators, and students. It also highlights the various potential area of healthcare where data science professionals, such as scientists, data engineers, and developers, can take the lead in building products and services that use Intel® technologies.
The human genome is a complex sequence of nucleic acids. It encodes as DNA within 23 chromosomes.6 It is well known that the expression of genes changes according to the situation and consequently such changes regulate many biological functions. Interestingly, certain genes change only as a result of specific pathological conditions (like cancer) or with treatment. These genes are called biomarker(s) for a specific tumor. Recently, a group of scientists from Oregon State University used a deep learning approach to identify certain genes critical for the diagnosis of breast cancer. As shown in Figure 2, they used a stacked denoising autoencoder (SDAE) for features extraction and then implied supervised classification models to verify new features in cancer detection.7 Another group of scientist from China applied a deep learning model for high-level features extraction between combinatorial SMP (somatic point mutations) and cancer types.8
Several endogenous molecules (chemical compound or protein) circulate in body fluid (blood, urine, cerebrospinal fluid). Some of these molecules are considered to be a tumor-specific biomarkers. The discovery of such a molecule or its synthetic analog gives new hope for understanding the mechanisms of disease and for creating therapeutic benefits.
The design of a new molecule is based on the historical dataset of old molecules and targets. In quantitative structure-activity relationship (QSAR) analysis scientists try to find out a known and novel patterns between structures and activity. At the Merck Research Laboratory, Ma et al. used a dataset of thousands of compounds (~5000) and built a model based on the architecture of DDNs.9 In another QSAR study, Dahl et al. built neural network models on 19 datasets of 2000‒14000 compounds to predict the activity of new compound.10 Aliper and colleagues built a deep neural network–support vector machine (DNN-SVM) model that was trained on a large transcriptional response dataset and classified various drugs into therapeutic categories.11
AtomNet is the first structure-based deep convolutional neural network. It incorporates structural target information and consequently predicts the bioactivity of small molecules. This application worked successfully to predict new active molecules for targets with no previously known modulators.12 Furthermore, Altae-Tran et al. introduced a new deep learning architecture called iterative refinement long short-term memory (LSTM), which significantly increases predictive power for specific drug discovery problems even with limited data.13
Histopathological images are primarily obtained from the thin section of a tumor. These sections are generally stained with specific colorful chemicals or antibodies to distinguish cancerous cells. Recently Kaggle* organized the Intel and MobileODT Cervical Cancer Screening competition to improve the precision and accuracy of cervical cancer screening using deep learning.14 The participants used different deep learning models such as the faster R-CNN detection framework with VGG16,15 supervised semantic-preserving deep hashing (SSDH), and U-Net for convolutional networks.16 As shown in Figure 3, Dr. Silva achieved 81 percent accuracy using the Intel® Deep Learning SDK and GoogLeNet* using Caffe* on the validation test.
Turkki et al. also applied convolutional neural networks (CNNs) and a support vector analysis approach to quantify immune cells in breast cancer slides. They achieved up to 90 percent of the agreement, similar to what pathologists achieve.17 Another application of deep learning is predicting the prognosis of cancer (that is estimating the stage of cancer). Hyung et al. found 83.5 percent average accuracy in predicting a patient’s survival with gastric cancer.18
Medical technologies such as computed tomography, magnetic resonance imaging (MRI), and ultrasound are a rich source to capture tumor images without invasion. Deep learning models can be used to measure the tumor growth over time in cancer patients on medication. As shown in Figure 4, Jaeger et al. applied CNN architecture on diffusion-weighted MRI. Based on an estimation of the properties of the tumor tissue, this architecture reduced false-positive findings and thereby decreased the number of unnecessary invasive biopsies. The researchers noticed that deep learning reduced the motion and vision error and thus provided more stable results in comparison to manual segmentation.19 A study conducted in China showed that deep learning helped to achieve 93 percent accuracy in distinguishing malignant and benign cancer on the elastogram of ultrasound shear-wave elastography of 200 patients.2,20
Several participants in the Kaggle competition successfully applied DNN to the breast cancer dataset obtained from the University of Wisconsin. Based on the features of each cell nucleus (radius, texture, perimeter, area, smoothness, compactness, concavity, symmetry, and fractal dimension), a DNN classifier was built to predict breast cancer type (malignant or benign) (Kaggle: Breast Cancer Diagnosis Wisconsin). Similarly, Xu et al. investigated datasets from over 7,000 images of single red blood cells (RBCs) from eight patients with sickle cell disease and applied the DNN classifier to classify the different RBC types (sickle, elongated, granular, oval, discocytes, and so on) (see Figure 5). The trained deep CNN distinguished the subtle differences in texture alteration inside the oxygenated and deoxygenated RBCs.22
Healthcare generates a huge amount of data, which is consistently growing. However, this data is not widely available to scientists and developers in startups and other areas of industry and in academic areas. While the potential for lucrative profit that can come from joining AI with the healthcare domain is high, unfortunately only a small fraction of startups and institutes are interested in exploring AI tools in healthcare. The amount of investment of time, money, and human resources in this domain is restricted probably due to ethical concerns and several rules and regulations. However the data scarcity trend is changing rapidly, and there is a lot of room in which to grow.
In order to build a deep learning model in the medical field, we need a significant amount of high-quality data. In some cases, the specific clinical or research data are not available (or are very limited) from a particular institute. It is essential to collaborate with other institutes to get sufficient data. But setting up such collaborations with a mutual agreement is sometimes hard to accomplish and very time-consuming.
People working in the medical, paramedical, and research fields are well educated for performing medical and research jobs. However, they may not have enough training in computer science and various computation languages, such as C++, Python*, and Java*, hardware and software knowledge. Having a team of AI experts, such as deep learning data scientists, developers, and solution architects, may be beneficial in order to fill the gaps in computation skill and run the healthcare projects.
It is essential that a data scientist, developer, or data engineer have knowledge of the healthcare domain and also a good understanding of DNNs as well as experience in advanced statistical modeling. They should be aware of the latest deep learning frameworks, libraries, APIs, UIs, web interactive notebooks, labs, Intel® AI DevCloud, multinode systems, big data, and so on. Implementations using older libraries or frameworks in the project may delay the process or produce poor results. A team of deep learning experts may help you to design optimal sampling procedures and get effective results. They may also help you to not only identify external and internal factors causing variation in your data analysis but also build strategies to reduce optimization difficulties due to poor-conditioning or local minima.23
Deep learning has a great potential to help medical and paramedical practitioners24 by
The histopathological examination of thousands of images is complex, time-consuming, and labor intensive. How can AI help?
A team from Harvard Medical School’s Beth Israel Deaconess Medical Center noticed a 2.9 percent error rate with the AI model and a 3.5 percent error rate with pathologists for breast cancer diagnosis. Interestingly, the pairing of “deep learning with pathologist” showed a 0.5 percent error rate, which is an 85 percent drop.24 Litjens et al. suggest that deep learning holds great promise to improve the efficacy of prostate cancer diagnosis and breast cancer staging. 25,26
Intel commends the AI developers who contribute their time and talent to help improve diagnosis and treatment for this life-threatening disease. Committed to helping scale AI solutions through the developer community, Intel makes available free AI training and tools through the Intel® AI Developer Program.
Intel recently published the series of AI hands-on tutorials. Here you will learn:
The same concepts could also be useful in healthcare to solve a similar set of problems. In the next series of articles, we will explore some examples of healthcare datasets where you will learn how to apply deep learning. If you want to test the deep learning in your own dataset, please contact the support community at Intel® AI Developer Program. Our Intel team will help you achieve your project goals.
Take part as AI drives the next big wave of computing, delivering solutions that create, use and analyze the massive amounts of data that are generated every minute.
Sign up to get the latest tools, optimized frameworks, and training for AI, machine learning, and deep learning.
Intel's compilers may or may not optimize to the same degree for non-Intel microprocessors for optimizations that are not unique to Intel microprocessors. These optimizations include SSE2, SSE3, and SSSE3 instruction sets and other optimizations. Intel does not guarantee the availability, functionality, or effectiveness of any optimization on microprocessors not manufactured by Intel. Microprocessor-dependent optimizations in this product are intended for use with Intel microprocessors. Certain optimizations not specific to Intel microarchitecture are reserved for Intel microprocessors. Please refer to the applicable product User and Reference Guides for more information regarding the specific instruction sets covered by this notice.
Notice revision #20110804