Deep Learning — What’s the hype about?
Healthcare Part 1 — Medical Diagnoses
To say artificial intelligence (AI) is transforming healthcare would be an understatement. Thanks to enormous advancements in computer processing power, as well as the increase of data collection at the patient, clinician and institutional level, AI is now driving the digital healthcare revolution. This transformation has been visibly apparent within the fields of medical diagnoses, drug discovery, e-health, and electronic health records. In particular, the field of medical diagnoses is where AI and deep learning have shown the most use cases. This is largely due to the recent developments in computer vision and object recognition, especially from 2012 when AlexNet, a convolutional network, won the ImageNet Large Scale Visual Recognition Challenge. In fact, the growth of AI research in healthcare and the advances in computer vision link closely together (see Figure 1). If implemented wisely, deep learning can continue to revolutionise healthcare, making it easier to access and more precise, thus resulting in better treatment outcomes for the patient.
Diagnostic Imaging
Naturally, medicine is divided into numerous specialties, therefore many different techniques are used to help diagnose illnesses. Diagnostic imaging is one such technique which is used across multiple medical specialties, and in which has been improving thanks to deep learning. Diagnostic imaging includes techniques like magnetic resonance imaging (MRI), computed tomography (CT) and positron emission tomography (PET), as well as different types of microscopy. Diagnostic imaging requires the doctor to use their training and experience to visually identify an irregularity (e.g. presence of a certain cell; irregular shape; colour etc). Here, the doctor’s role resembles a typical classification task that neural networks are adept at solving, thus highlighting the potential of deep learning in medical diagnoses. Below are just some of the growing number of uses cases within the field.
Alzheimer’s Disease
Within ageing and dementia research over the past decade, numerous studies (Liu et al., 2015; Lu, Popuri, Ding, Balachandar, & Beg, 2018; Suk, Lee, Shen, & Alzheimer’s Disease Neuroimaging, 2014) have been published discussing the use of ANNs, particularly CNNs, to analyse brain scans (generally MRIs) in order to classify and distinguish Alzheimer’s Disease (AD) from healthy controls and age-related decline. AI has the potential to revolutionise early detection and treatment, and thus could reduce the rates of AD and dementia, which was recently identified as the leading cause of death in the UK. Early detection has proven difficult with neurodegenerative disorders such as dementia and AD, as neurobiological changes (e.g. brain changes visible on MRI) often precede clinical symptoms (e.g. subjective memory complaint) by several years. Furthermore, given the enormous variance of brain structure among the population, subtle neurobiological changes are likely too difficult for the clinician to identify.
Recently, CNNs have been used to predict early-stage AD using 18-F-fluorodeoxyglucose PET (FDG-PET) imaging. Using the Alzheimer’s Disease Neuroimaging Initiative dataset (2109 scans; 1002 patients), Ding et al. (2018) trained an InceptionV3 CNN on 90% of the dataset and then tested it on the remaining 10%. The algorithm was also tested on a separate dataset of 40 new scans, some of which were later clinically diagnosed with AD. The CNN was able to predict AD an average of 6 years before clinical diagnosis at 100% sensitivity (true positive rate), compared to 57% sensitivity by radiologists. This is just one of a number of studies highlighting how deep learning could help early detection in dementia and AD.
Oncology
Cancer is the second leading cause of death worldwide and was estimated to cost the global economy US$1.16 trillion in 2010. Generally, research has been targeted towards prevention and treatment; with less focus on early detection and diagnosis. ANNs have been successfully used to aid in classification and segmentation, which could lead to better health outcomes for the patient. Published in Nature, Esteva et al. (2017) illustrated that a deep CNN trained on 129,450 clinical images of 2,032 different diseases was able to identify the most common form of skin cancer (keratinocyte carcinomas) and the most deadliest form of skin cancer (malignant melanoma) from benign lesions with the same accuracy as 21 dermatologists. Findings like these illustrate that implementing deep learning could result in greater time-efficiency for doctors and quicker results for patients. Furthermore, the authors suggest that with the right technology, a smart phone could be become a powerful diagnostic tool; a ‘selfie’ could be all that is needed to identify whether an irregular mole warrants medical intervention.
CNNs have also been extensively used in brain tumour segmentation. The success of radiation and chemotherapy rely heavily upon differentiating or segmenting the tumour from healthy tissue. This is exceptionally difficult for brain tumours like gliomas and glioblastomas, where shape, size, location, as well as relative contrast and diffusivity to healthy tissue have high inter-patient variance. To improve segmentation, Havaei et al. (2017) proposed a neural network that emphasises local dependencies, which they claim traditional pixel classification often overlooks. Furthermore, their final layer was a convolutional implementation of a fully connected layer, which allowed for a dramatic increase in speed. In another study, Pereira, Pinto, Alves, and Silva (2016) used 3x3 kernels which allowed for a deeper network and resulted in less overfitting. Furthermore, to reduce heterogeneity caused by multi-site and multi-modal data acquisition, they used intensity normalization during pre-processing, which although uncommon, improved segmentation.
Arrythmia
Arrythmias refer to a change in the electrical impulses of the heart, and are often categorised as a slow, fast or irregular heartbeat. Although presence of an arrhythmia doesn’t always warrant intervention, if paired with other symptoms, it could be an early indicator (risk factor) for stroke, heart failure or cardiac arrest. Given that cardiovascular diseases are the leading cause of death globally; early detection is vital. Currently, arrythmias are detected via an electrocardiogram (ECG), a device that graphs the heart’s electrical activity. Traditionally these devices require 12 leads to be placed in specific locations on the patient, but now measurements can be taken and interpreted using wearables such as smart watches (e.g. Apple Watch Series 4), in a method known as single lead ECG. Whilst automated ECG interpretation has been around since the 1970’s, the use of deep learning in recent years have made them far more reliable. CNNs are particularly useful, with one study illustrating their 34-layer CNN could detect a wide range of arrythmias more accurately from a single-lead ECG than a group of board-certified cardiologists (Rajpurkar, Hannun, Haghpanahi, Bourn, & Y. Ng, 2017). Also using single lead ECG, another study trained a 16 layer deep 1D-CNN to classify 17 types of arrythmias with an accuracy rate of 91.33% (Yıldırım, Pławiak, Tan, & Acharya, 2018). Furthermore, the authors employed a unique 1D max pooling layer, reducing the necessary computational power, thus making the algorithm suitable for smart phones and e-health.
Pathology
Pathology is the study of diseases using classical laboratory techniques with specimen samples (e.g. blood, urine and tissue). Analyses in hemopathology (blood) and histopathology (tissue), involve several specific methods, however most eventuate in the creation of gigabytes of data which the pathologist visualises and examines through a microscope and related software packages. As such, AI is proving to be rather valuable in pathology, especially regarding the counting of events (e.g. labelled cells), segmentation of cells (e.g. nuclei) and the classification of tissues (e.g. healthy or diseased). Deep learning algorithms can implicitly learn to recognise patterns making them superior to the current state of the art hand-crafted feature-based approaches. Although the latter uses elements of machine learning, their requirement of more regular pathologist intervention make them less efficient. Similar to imaging diagnosis, CNNs appear to be the most suitable for pathology. More detailed use cases are outlined in Janowczyk and Madabhushi (2016).
The interest of AI in the field of diagnostic imaging is further evident with the recent release of NVIDIA CLARA, an open platform which promises to develop next generation medical instruments. The platform uses deep learning algorithms to sort through vast amounts of imaging/medical data, so it can be useable by clinicians and doctors to assist in early detection, diagnosis and treatment of diseases.
Genomics
A genome is the complete set of genetic material present in an organism; genomics is the study of genomes. Since the late 1990’s researchers have been gathering masses of genomic data to better understand how our genomes interact with the environment. With improvements in genome sequencing and analytical methods, researchers hope to better understand the underlying mechanisms of genes, and thus why and how specific diseases develop in certain individuals. In recent years, a variety of ANNs have been implemented within genomics. Given their outstanding feature extraction ability in computer vision, if a section of the genome sequence is captured as an image, CNNs can be used to discover meaningful patterns and thus classify the genomic data. Genomes or more so DNA is extremely sequential, which means that RNNs, which have been proven to be well adept in other sequential data like natural language processing, can be rather useful. Some use cases of RNNs have been to quantify the function of DNA sequences, and to convert signals (from sequencing instruments) into nucleotide sequences. Uses cases of ANNs within genomics are outlined in more depth in Yue and Wang (2018). Large tech companies are beginning to see the value deep learning within genomics, with IBM recently launching IBM Watson for Genomics, an initiative which aims to provide comprehensive, validated, and up-to-date insights to aid clinical reporting in oncology.
Chatbots
The above use cases show how deep learning can help clinicians to improve the outcomes of their patients, but patients can also benefit directly. Chatbots, as the name suggests, are messaging platforms where the user converses with an AI avatar designed to respond like a human. Within healthcare, chatbots are used to answer health related questions. For example, a patient may type in symptoms they are experiencing, and based on follow-up questions, the patient’s history, and the AI’s clinical database, the chatbot will provide health information and potential recommendations. This is a rather serious improvement in healthcare, as empowering patients to actively monitor their own health is key to ensuring long-lasting health benefits. Chatbots use a combination of machine learning and deep learning algorithms. Given chatbots in healthcare are relatively new, constantly developing and a potentially highly lucrative industry, few companies discuss their algorithms or neural networks. However, given that chatbots are designed to potentially classify and/or identify information based on text input, one can assume they are heavily dependent on algorithms used within natural language processing and classification e.g. FFNs, CNNs and RNNs. Some companies currently paving the way include Babylon Health, Your.Md and Ada Health.
As you can see, deep learning is already having a significant impact in healthcare, specifically within medical diagnoses. For most of the uses cases outlined above, the next step is large-scale implementation, which will no doubt come with time.
In the next post, we will continue outlining deep learning in healthcare, with a focus on drug discovery, e-health, and electronic health records.
References
Ding, Y., Sohn, J. H., Kawczynski, M. G., Trivedi, H., Harnish, R., Jenkins, N. W., . . . Franc, B. L. (2018). A Deep Learning Model to Predict a Diagnosis of Alzheimer Disease by Using 18F-FDG PET of the Brain. Radiology, 0(0), 180958. doi:10.1148/radiol.2018180958
Esteva, A., Kuprel, B., Novoa, R. A., Ko, J., Swetter, S. M., Blau, H. M., & Thrun, S. (2017). Dermatologist-level classification of skin cancer with deep neural networks. Nature, 542, 115. doi:10.1038/nature21056
Havaei, M., Davy, A., Warde-Farley, D., Biard, A., Courville, A., Bengio, Y., . . . Larochelle, H. (2017). Brain tumor segmentation with Deep Neural Networks. Medical Image Analysis, 35, 18–31. doi:https://doi.org/10.1016/j.media.2016.05.004
Janowczyk, A., & Madabhushi, A. (2016). Deep learning for digital pathology image analysis: A comprehensive tutorial with selected use cases. J Pathol Inform, 7, 29. doi:10.4103/2153–3539.186902
Liu, S., Liu, S., Cai, W., Che, H., Pujol, S., Kikinis, R., . . . Fulham, M. J. (2015). Multimodal neuroimaging feature learning for multiclass diagnosis of Alzheimer’s disease. IEEE Trans Biomed Eng, 62(4), 1132–1140. doi:10.1109/tbme.2014.2372011
Lu, D., Popuri, K., Ding, G. W., Balachandar, R., & Beg, M. F. (2018). Multimodal and Multiscale Deep Neural Networks for the Early Diagnosis of Alzheimer’s Disease using structural MR and FDG-PET images. Sci Rep, 8(1), 5697. doi:10.1038/s41598–018–22871-z
Pereira, S., Pinto, A., Alves, V., & Silva, C. A. (2016). Brain Tumor Segmentation Using Convolutional Neural Networks in MRI Images. IEEE Transactions on Medical Imaging, 35(5), 1240–1251. doi:10.1109/TMI.2016.2538465
Rajpurkar, P., Hannun, A., Haghpanahi, M., Bourn, C., & Y. Ng, A. (2017). Cardiologist-Level Arrhythmia Detection with Convolutional Neural Networks. arXiv. doi:https://arxiv.org/abs/1707.01836
Suk, H.-I., Lee, S.-W., Shen, D., & Alzheimer’s Disease Neuroimaging, I. (2014). Hierarchical feature representation and multimodal fusion with deep learning for AD/MCI diagnosis. NeuroImage, 101, 569–582. doi:10.1016/j.neuroimage.2014.06.077
Yıldırım, Ö., Pławiak, P., Tan, R.-S., & Acharya, U. R. (2018). Arrhythmia detection using deep convolutional neural network with long duration ECG signals. Computers in Biology and Medicine, 102, 411–420. doi:https://doi.org/10.1016/j.compbiomed.2018.09.009
Yue, T., & Wang, H. (2018). Deep Learning for Genomics: A Concise Overview. arXiv.
