This installment of AI in Healthcare highlights the use of CNNs in the field of cancer tumour diagnosis. The AI in Healthcare series covers short updates delving into recent innovations, future directions, and patentability issues in the area of patient support technologies, telehealth, research, and clinician support technologies.
It has been 35 years since Yann LeCun, Chief AI Scientist at Meta, introduced the LeNet convolutional neural network (CNN). This CNN is most widely known its practical use in helping the United States Postal Service identify handwritten numbers, at a time when this was considered a technological breakthrough. Today, much of the recent buzz surrounding neural networks is for their use as a generative tool for creating images, text, and audio. However, the more traditional use of neural networks, and particularly CNNs, as classifiers still has much to offer in healthcare, even 30 years on.
To illustrate the application of neural networks and their use for diagnosing medical conditions, a specific application of CNNs to the detection of cancer tumours is our focus for this installment. We will cover what a CNN is, how it can be applied to the detection of cancer tumours, and what to look out for if you are looking to patent inventions in this space.
What is a CNN?
Like all other neural networks, CNNs involve processing a series of inputs through ‘layers’ of neurons. Specifically, neurons in one layer receive input from neurons in the previous layer and provide an output to neurons in the next layer. What distinguishes CNNs is the concept of a ‘receptive field’. The receptive field refers to a region of the input image, which determines which neurons get activated in response to an input within the receptive field. The neurons are organised in layers, that detect certain features, for example a neuron in one layer will react to an edge in its receptive field and a neuron in another layer may react to a corner in its receptive field. The neurons feed their activations from one layer to the next, for example ending up in a layer of face neurons that react to faces. The use of this concept was inspired by how real neural networks work in the brain, specifically the visual cortex, where you can find cells that specialise in detecting a certain feature (say an edge, corner, or face) in a specific location of the visual field known in neuroscience as that cell’s receptive field. The really interesting part is that the CNN defines the structure of receptive fields and layers but the features (edges, corners, faces) are not hard-coded but learned from the data as an efficient representation of the input to result an arrangement resembling the visual cortex of primates.
By incorporating this concept into a neural network, the CNN can take advantage of the inherent spatial relationship between a group of neighbouring pixels, without needing to process input from every pixel all at once. In this way, the CNN learns from images to build up a generalised representation of the image in terms of learned features akin to the features found in the visual cortex. This aspect combined with ‘weight sharing’, tying neurons in a layer across receptive fields (to make them detect the same features), greatly improves the scalability, efficiency, and classification performance of CNN as compared to less structured types of neural network. A true example of brain-inspired design leading to an efficient neural network architecture.
How can CNNs be applied?
The diagnosis of most cancers typically involves working with images of some sort, whether it be images taken from an MRI, CT, or PET scan, or histology images obtained from a biopsy. From these images, a radiologist can examine the images to diagnose a condition that may be present. Like any other neural network model, the CNN must be trained before it can be deployed for inference. In these kinds of use-cases, training neural networks involves obtaining images that have been labelled by experts. In this case, the training data usually comprises processed versions of medical images and corresponding labels indicating a diagnosis as applied by radiologists who have studied the images.
CNNs can be used to identify and classify various aspects of cancer tumours other than detecting their mere yes/no presence in an input image. For example, they can also be used to identify the cancer type and its severity or stage, as well. Depending on what the CNN is trained to classify, its architecture and training data can be adjusted accordingly.
For example, in “Deep learning ensemble 2D CNN approach towards the detection of lung cancer” by Shah et al. published in Nature in 20231, an ensemble approach was developed which adapts existing CNN architecture to more accurately detect early signs of lung cancer. As noted in the paper, it can be difficult to differentiate between lung nodules (a growth of abnormal tissue) and healthy lung tissue, leading to late diagnosis of lung cancer. In their paper, an ensemble approach was presented which combines three CNN models to work together, achieving an impressive accuracy of 95%, which is the record for accuracy of any deep learning algorithm for identifying lung cancer to date. This shows exciting potential of neural networks which can be used to identify different aspects of cancer progression.
How does this relate to patents?
The kind of image processing discussed above is something the EPO considers to be technical, so that many of the complexities that can occur when seeking to protect AI innovation with a European patent fall away. Nevertheless, it is important to focus on the “technical” features relating to how the images and data are processed by a computer. If the definition of the invention critically relies on the quality of the radiologists labelling, rather than on how the algorithm operates, trouble lies ahead and the EPO may consider that any advantage achieved by the invention relies on radiologist’s intellectual contribution so that the invention may be obvious for lack of a technical effect, as the EPO found in an appeal decision discussed in this article. Although image processing is one of the most well-established field for algorithmic patentability, it is crucial to focus on how an advantage is achieved by virtue of the processing being done by the computer.
Closing remarks
For further information or advice related to patenting inventions in this area, please do not hesitate to contact Alexander Korenberg, Nick Noble, or Aaron Yat Lam, or your usual Kilburn & Strode advisor.
If you enjoy our series and wish to discuss any of the related content or are interested in protecting your own innovation in these areas, do reach out to our Bioinformatics and Digital Health team.
1. Shah, A.A., Malik, H.A.M., Muhammad, A. et al. Deep learning ensemble 2D CNN approach towards the detection of lung cancer. Sci Rep 13, 2987 (2023). https://doi.org/10.1038/s41598-023-29656-z