Diagnosing Cancer with Machine Learning

Colorectal cancer is the second largest cause of cancer death in America, with an estimated 95,520 new cases of colon cancer and 39,910 new cases of rectal cancer in 2017. The large majority of colon cancers begins as a growth called a polyp, on the inner lining of the colon. There are two main types of polyps - adenomatous polyps and hyperplastic polyps. Adenomatous polyps sometimes change into malignant tumors. Therefore, adenomas are called a pre-cancerous condition.

Currently, the gold standard for detection is the colonoscopy. A doctor inserts a colonoscope into the rectum of a patient and manipulates it to the beginning of the stomach, at the end of the intestine. Then, the doctor begins withdrawal and examination of the colon. This process can take 8-20 minutes, depending on the circumstances. A colonoscopy procedure requires a doctor to visually examine the colon for polyps. In a tandem colonoscopy, the differences in the number of adenomas found was as much as 27%.1 In a study supported by The National R&D Program for Cancer Control, Ministry of Health & Welfare, Republic of Korea, researchers concluded that “there was a significant miss rate in the detection of colonic adenomas even in quality-adjusted, back-to-back colonoscopies” and “ the development of specific technological methods to reduce the adenoma miss rate is necessary.” It has been estimated that up to 50% of new cases are caused by polyps missed during colonoscopy.2

I am researching how to use computer vision and machine learning to detect polyps. Explicitly describing our processes for object recognition is virtually impossible as each case must be accounted for to detect objects in those instances. Computer vision is “concerned with the automatic extraction, analysis, and understanding of useful information from a single image or a sequence of images”, as defined by the British Machine Vision Group.3 Artificial intelligence is defined as the “field of study which studies how to create computing systems that are capable of intelligent behavior”. Machine learning is a branch of artificial intelligence. One of the goals of this field is to teach a computer to “learn” without giving it explicit directions.

Deep learning is a branch of machine learning that uses algorithms, sometimes inspired by the brain, to represent data. Deep learning usually involves neural networks, which is essentially a big function, with many layers - each with their own weights and biases that affect the output to each successive layer of neurons. The weights of each neuron are adjusted using a technique called backpropagation. Backpropagation is an algorithm first proposed in 1986 by Geoff Hinton4 that uses the loss function (found by comparing the output values of the neural network function to those of the ground truth) of the neural network to make adjustments to the relative weights and biases of neurons in the hidden layers to minimize the output of the loss function.

My project is essentially an object detection, supervised learning problem. Convolutional neural networks (CNN) have been incorporated into almost every image recognition problem. They possess several advantages over standard neural networks, because “ConvNet architectures make the explicit assumption that the inputs are images” which allows researchers to “vastly reduce the amount of parameters in the network,” as stated in Stanford’s CS231n course.5 CNNs differ from regular neural networks in that they arrange their neurons in three dimensions according to height, width, and depth. They generally have three types of layers: the convolutional layer, the pooling layer, and the fully connected layer. Developing a network with the ability to classify images that contain polyps and draw bounding boxes around polyps in real time is the end goal of this project. Machine learning has many pertinent applications and a huge potential to improve quality of life.