How to Deploy an H2O.ai Machine Learning Model to AWS SageMaker
Artificial Intelligence Development
In the tech world, there is no shortage of trendy buzz words that server to appetize, excite and ultimately disappoint the masses once their PR-fueled sheen fades. However, Deep Learning and its applications to medical imaging, is one technology where the hype is completely deserved. We are on the precipice of a revolution in radiology as the use of Deep Learning in is set to transform how medical imaging is done and drastically improve patient outcomes by making diagnoses that are quicker, accessible and as accurate as those made by radiologists today. According to Signify Research, it is estimated by 2023 machine learning powered medical imagining will have grown to a $2 billion in annual revenue.
At Blue Label Labs, we work on a number of new and emerging technologies and none excites us more than the potential of artificial intelligence (AI) especially in the form of machine learning and deep learning. The technology is exciting to behold, however we believe that there is tremendous near-term opportunity for innovation using machine and deep learning within medicine, specifically in medical imaging and radiology.
A Quick Overview on Deep Learning and Machine Learning
Machine learning (ML) is a blanket term used to describe a field of computing, of which Deep Learning is a subset, by which computers can learn to recognize patterns in data without being explicitly programmed to do so. Normally, a computer algorithm is a step by step sequence of statements, conditionals and loop operators that must be declared by the programmer as explicit instructions. Machine learning is a technique by which a computer is able to ‘learn’ how to solve a problem by essentially learning through past experience. As a simple (and naive) example, we can develop an ML model to predict the likelihood of someone contracting heart disease based on a number of inputs such as age, ethnicity, BMI, etc. An ML model to solve this problem isn’t created by a programmer writing lines of code, instead a set of “training” set of data is fed to the computer, in this case probably a large spreadsheet of past patient medical data, where each row of that training data contains a real person’s age, ethnicity and BMI, along with some indication if they contracted heart disease or not. The computer analyzes all this data and tries to generalize the relationship between the inputs (age, ethnicity, BMI) and the outputs (did the person contract heart disease?) in the form of a model or equation. For those of you who remember high school math, one approach to generalizing a series of data points is through linear regression, which is precisely one type of machine learning approach. Once the model has been trained, the equation that the computer has deduced from its training data can then be used to make predictions on the likelihood of a patient developing heart disease based on their age, ethnicity and BMI. While ML is now a popular buzzword thrown around the industry, it is by no means new and these techniques have been in use for decades. However, traditional machine learning techniques are generally limited to problems where a linear relationship can be made between the inputs to the algorithm and the outputs. This limitation often made machine learning inapplicable to a great deal of modern problems, such as image recognition where the relationship is far from linear and much more complex.
Deep Learning is a subset of Machine Learning and uses more sophisticated constructs to solve problems where a simple linear relationship cannot be generalized between inputs and outputs. In Deep Learning, more complex structures such as multi-layered Convultional Neural Networks are used to help identify relationships between inputs and outputs that often require multiple, non-linear transformations. A great application of Deep Learning is in image recognition, where trained models using CNNs are now able to achieve an accuracy of 98% in image classification of the ImageNet data set compared to 86% in 2012. The rapid increase in computing power, the availability of large labelled datasets and the development of easy-to-access Deep Learning frameworks such as TensorFlow and PyTorch have liberated the practical application of Deep Learning techniques from academia and into private industry.
Applications of Deep Learning to Medical Imaging
The possibilities in deep learning are great and the possible applications to medicine are exciting, however the area of medicine set to be revolutionized by these machine learning techniques is in medical imaging. Imagine a mobile app that can take a CT scan as an input and can identify a tumor in that person’s brain. This is not some far fetched vision of the future, this type of innovation is happening right now! Currently medical imaging is a manual process which relies on highly-trained human judgement to perform this analysis, which is slow and can be prone to errors in human judgement. Furthermore, the labor pool of trained radiologists to make these decisions is limited, making it difficult for people in underdeveloped countries or in remote locations to be quickly diagnosed. Given that it is estimated that 90 percent of all medical data today is in the form of images, the power of deep learning, especially within the problem domain of image recognition, is poised to change how medical imaging and diagnosis is done for a number of medical subspecialties:
Melanoma Detection
IBM acquired Merge Healthcare for a cool $1 billion in 2015, however it might surprise you that the IBM didn’t cut a billion dollar check to get access to some bleeding edge algorithm, but rather to get access to the 30 billion images that Merge Healthcare owns. IBM plans for this acquisition were to use relevant medical imagery owned by Merge Healthcare to develop deep learning models for the detection of melanoma and other tumors. The results of this acquisition have already appeared to bear fruit during the 2016 International Skin Imaging Collaboration (ISIC) International Symposium on Biomedical Imaging challenge, which compared the performance of automated deep learning models with that of dermatologists in the diagnosis of melanoma through medical imagine. The results of this challenge showed while on average the deep learning algorithms performed on par with that of dermatologists, it often out-performed dermatologists in the diagnosis of melanomas.
Stroke Detection
Neurologists will tell you that the prompt and accurate identification of a Large Vessel Occlusion (LVO) is key to the long term survival and recovery of a patient suffering an Acute Ischemic Stroke (AIS). Deep learning techniques have already come to the market to help neurologists identify in medical imaging the onset of stroke within a patient. In 2018, the FDA approved a CNN-based algorithm developed by Viz.AI that is able to rapidly detect an occurrence of LVO within a patient’s CT image. Viz.AI is a mobile app backed by a CNN model that processes uploaded CT images to diagnose instances of LVO. In a study submitted by the company, over an analysis of 650 images, the Viz.AI deep learning algorithm demonstrated a sensitivity (true positive detection rate) of 82% and a specificity (true negative detection rate) of 94% with a median scan time of 6 minutes, eliminating on average 52 minutes in the typical time taken to manually analyze these images. While not perfect, the Viz.AI algorithm demonstrates that deep learning models are good enough to be applied in clinical settings and make measurable positive impact to patient recovery.
Diabetic Retinopathy
Diabetic Retinopathy is a complication of diabetes that is caused by damage of the blood vessels in the retina causing blindness. Screening of diabetic retinopathy is done through imaging of the retinas and early detection is key for a good prognosis. A variety of scientific studies have been done to demonstrate how Deep Learning techniques can be used to develop an automated system for the detection of Diabetic Retinopathy with very high sensitivity. In a study published in Nature, a CNN powered detection model was demonstrated to achieve a sensitivity of 90% and a specificity of 97%. The use of Deep Learning in the diagnosis of diabetic retinopathy has been shown to hold great promise by this study, however, unlike stroke detection, no private company has come to the market with a practical application of this technology as of yet.
Opportunity for Disruption via Deep Learning
Deep learning is set to revolutionize medical imaging and radiology, which will ultimately lead to better patient outcomes. However, it also represents a tremendous green field opportunity for medical innovators to reap a financial windfall in the practical application of deep learning techniques in the field. While companies such as Viz.AI have brought to market successful application of deep learning within particular fields of medical imagery, there exists tremendous opportunities in areas such as the diagnosis of medical retinopathy for startups to bring to market techniques already shown to be promising in academic studies in the form of monetizable products and services. The time is ripe for the practical application of Deep Learning in medical imaging and radiology and the industry awaits who the ultimate champion of this technology might be.
Is your company looking to take the plunge into machine learning and deep learning based products? Our team at Blue Label Labs has a deep bench of expertise in both product design and development, but also the application of machine learning in the solving image recognition problems. Reach out to one of our experts to see how we can partner to bring your product to life.
Bobby Gill
