The current role that Artificial Intelligence (“AI”) plays in society can be summarised by a single quote — “AI is the new electricity.”1 According to the World Intellectual Property Organization, AI patents are already being filed for applications in some of the world’s largest industries such as transportation, telecommunications, education, and business. In particular, the life and medical sciences industry is mentioned in 19% of all AI-related patent documents,2 with most of these patents utilising machine learning or computer vision methodologies.3 Consequently, medical imaging has become a significant area of application for AI, representing the second-highest number of patent filings within the life and medical sciences industry.4
AI algorithms in medical imaging generally use deep learning architectures that employ Convolutional Neural Networks (“CNN”). This is unsurprising, as CNNs were designed to map image data to an output variable and are the predominant method used for prediction problems that involve image data as the input.5 Additionally, CNNs are able to develop internal representations of 2-D images, which allow the model to learn features in an image in a manner that is less susceptible to variance in position and scale compared to alternative machine learning architectures.6 The ability to learn these features becomes significant in the context of medical images, as it allows a CNN to extract features of complex pathologies that may be present in an image, and differentiate it from images where such pathologies are not present.
The role AI plays in the medical imaging space is already apparent, with many active companies addressing areas such as breast imaging, cardiovascular imaging, and diagnostic imaging.7 Furthermore, AI has shown great clinical efficacy in research settings. For example, Google’s Lymph Node Assistant, or LYNA, has been shown to correctly distinguish between pathology slides of a lymph node with and without metastatic cancer 99% of the time.8 Additionally, an AI system developed by DeepMind has been shown to match or beat world-leading eye specialists by correctly referring patients with more than 50 different eye diseases for further treatment with a 94% accuracy.9 AI has also been shown to be effective in dermatology, with a system being able to match the performance of 21 dermatologists in determining whether skin lesions were cancerous.10
The examples above also have associated patents that have been filed. Google’s LYNA is described in P.C.T. Application No. PCT/US2017/019051, which outlines the use of CNNs to analyse images of lymph node tissue. DeepMind’s AI system is described in U.S. Patent No. 10,198,832, which highlights their use of segmentation neural networks for analysing Optical Coherence Tomography images of the eye. The patent states that the segmentation neural networks may be implemented to include CNN layers. Finally, U.S. Patent No. 10,223,788, assigned to IBM, describes the use of CNNs to predict whether a skin lesion is present in a dermoscopic image.
Although patents in the AI space may seem to be dominated by large companies such as Google, Siemens (which owns the highest number of patents in the medical imaging space11), and IBM (which owns the highest number of U.S. patents in all industries12), emerging startups have also filed patents for applications suited to their use of AI. One such startup is Subtle Medical, which has developed software for shortening imaging times using a Deep Residual Network (“ResNet”), an improvement on the traditional CNN.13 Their patent (U.S. Patent No. 10,096,109) describes a method of preprocessing medical images which are then input into a ResNet, thereby improving the quality of the images. Another startup, Butterfly Network, incorporates AI into their handheld Ultrasound device to guide users in positioning the Ultrasound device during its use, based on real-time analysis of the acquired Ultrasound images. Their technology is described in U.S. Publication No. 2019/0059851, which includes a method for instructing an operator to acquire Ultrasound images by moving the device along an area on the body, followed by a machine learning technique to identify whether the acquired images match a “target” anatomical view. This patent provides three examples of machine learning techniques, all of which utilise a CNN. Finally, Zebra Medical Vision, a startup with the goal of using deep learning to analyse medical scans for a dollar each, has developed a deep learning engine to automatically detect diseases present in medical images acquired from Computed Tomography, Magnetic Resonance Imaging, and Ultrasound scans. This deep learning engine is described in U.S. Publication No. 2018/0240235, which includes methods for automatic segmentation of medical images to identify anatomical features through the use of CNNs.
AI is proving to be an effective tool for use in the medical imaging field, with new research continuously showing its clinical efficacy in a variety of applications. Many companies have translated such research into industry opportunities and developed tools to be used for medical imaging applications that are based around AI-driven algorithms. Patenting such AI or machine learning technology for use in the medical imaging field, however, comes with some unique challenges. Simply processing and analyzing visual images may not qualify as eligible subject matter. Depending on the jurisdiction, the patent office may require that a specific technical problem is being overcome by the invention, or that the invention involves an element of physicality. As well, the enforceability of an AI patent, given the “black box” nature of many AI systems, also poses another challenge. Although CNNs are currently the predominant AI architecture used in the above patents, newer algorithms are also on the rise. Innovators and practitioners looking to protect AI imaging technology may have to think outside the black box when fitting AI technology, with all its unique properties, into each country’s patent system.