AI in Focus: Image Recognition
The next application of Artificial Intelligence (AI) to be examined in our AI in Focus series is in the area of image recognition, which typically involves determining regions of interest in an image and classifying those regions of interest.
Recent advances in computer vision algorithms, artificial intelligence (AI), and computing power have enabled significant growth in the field of image recognition. In comparison to early edge detection and segmentation techniques, modern image recognition is significantly faster, more accurate, and generally more powerful. In many cases, AI-based technologies can now analyze and process images at a level equal to, or even greater than, human visual systems. In fact, the recent Canadian Intellectual Property Office (CIPO) report on AI shows that AI patent activity for computer vision based inventions has been the fastest growing segment compared to other AI applications for the last 20 years.1
One (humorous) example of an AI image recognition system is the mobile application Not Hotdog, popularized by the HBO television series, Silicon Valley.2 Not Hotdog identifies whether (or not) an image includes a hotdog. While Not Hotdog may appear to be more whimsical than useful, it illustrates that AI technology has several characteristics that make it suitable for use in a number of meaningful image recognition based applications.
For example, some practical applications of AI image recognition include:
- inspection, e.g., for manufacturing and industrial processing;
- identification, e.g., for security or fraud prevention;
- object tracking, e.g., for logistics or surveillance;
- diagnostics, e.g., for medical imaging; and
- navigation, e.g., for autonomous vehicles3 and other robotics.
Medical imaging and diagnostics is a powerful application area for AI image recognition. AI image recognition technologies can aid medical professionals in the interpretation of medical images by detecting certain objects of interest and the diagnosis of disease by more accurately classifying the detected objects. Patent filings related to medical imaging and AI have recently risen significantly, up 680% since 2012. Some of the top applicants in terms of number of filings notably include Siemens, GE, IBM, Tencent, and Philips.
One example of a medical diagnostic patent that uses AI image recognition is U.S. Patent No. 10,025,902 entitled “Enhanced Pathology Diagnosis”, assigned to Verily Life Sciences LLC (a Google subsidiary).4
The patent is generally directed to methods for enhancing pathology diagnosis using a machine learning algorithm. The patent describes how a machine learning algorithm can be trained to alert pathologists to regions of interest in a pathology sample, such as diseased tissue, atypical cells, unusual growths, etc. The patent explains that the machine learning algorithm can be used to provide a second opinion to a pathologist, increasing pathologist efficiency, and decreasing the probability of an inaccurate diagnosis.
This patent also discloses how a microscope is used to record magnified pathology images of a pathology sample. A machine learning algorithm, trained using reference pathology images, compares the magnified pathology images to the reference pathology images to identify regions of interest in the magnified pathology images. A user is then alerted through audio or haptic notification of the identification of regions of interests and the provision of a second opinion. Notably, the patent does not describe a specific machine learning algorithm in detail, and instead lists a number of different possible approaches, including neural-networks, rule learning, deep learning, inductive logic programming, decision tree learning, etc.
Recent filing trends suggest that many more patent applications directed toward AI image recognition will be filed in the near future. Undoubtedly, many of these patent applications will relate to medical imaging and diagnostic applications.
Another factor that will influence the number of patent filings directed to AI-based image recognition are recent changes to how Canadian Intellectual Property Office (CIPO) assesses patentable subject matter in the field of computer-implemented inventions. A Practice Notice5 was released following the Federal Court decision in Yves Choueifaty v Attorney General of Canada 6. This CIPO Practice Notice states that in order to constitute an actual invention, a disembodied idea, scientific principle or abstract theorem must cooperate with a combination of elements to form a single invention and that the combination must have physical existence or manifest a discernible physical effect or change and relate to the manual or productive arts.
However, the mere fact that a computer is included in the claim is not a guarantee that the claim will be considered as having patentable subject-matter.7 The CIPO Practice Notice notes that if a computer merely processes a mathematical algorithm in a well-known manner, the computer and the algorithm do not form part of a single actual invention that solves a problem related to the manual or productive arts. On the other hand, if running the algorithm on the computer improves the functioning of the computer, then the computer and algorithm would together form a single actual invention and the subject-matter defined by the claim would be patentable subject matter.
In general, CIPO’s new approach to assessing patentable subject matter should be welcomed by patent applicants wishing to patent AI image recognition inventions in Canada whose inventions have a physical existence, manifest a discernible physical effect or change or serve to improve the functioning of computers.