Colorectal cancer treatment can become more efficient following two breakthroughs: a machine learning system that has dramatically improved detection and a web tool that can predict the risk of its recurrence.
In In 10 seconds? Researchers have achieved a 100% accuracy in predicting colorectal cancer using an advanced form of AI, known as machine learning. Independently, another team has used a risk model to create a tool that highlights patients whose cancer is likely to come back. (Read the science)
What does it mean for patients? The achievements can lead to better outcomes when rolled out in the clinic. The machine learning technique can improve early detection. It can distinguish precancerous polyps and healthy tissue with amazing accuracy. Current diagnosis is performed by doctors analysing images from an endoscope-mounted camera and can result in missing early warning signs. Also, this method can only spot changes in the surface of the bowel wall and not in deeper layers. (Read the preprint paper)
And how does improve diagnosis? Mounted on endoscopes, it will provide real-time, deeper tissue layer analysis and identify cancerous tissue with a 100% accuracy! This will be a great aid to clinicians diagnosing colorectal cancer as they will be able to zoom in and find precancerous polyps and early-stage cancers in deeper tissue regions. (More on OCT in colorectal cancer diagnosis)
How did they do it? Researchers noticed that healthy colorectal tissue had a teeth-like pattern, but it was rarely present in precancerous or cancerous tissue. They turned to Optical Coherence Tomography (OCT), a method that uses infrared light waves to capture 3D images of structures inside the tissue. Then they trained a neural network to learn structural patterns found in 26,000 images of patients’ tumour, benign and abnormal areas. The pattern recognition algorithm was able to pick out with a 100% accuracy tumour and healthy tissue! Which is great, because when found early, colorectal cancer is highly treatable. (More on economic implications of early detection)
And what about the recurrence-prediction tool? This is a web interface based on data from about 8,300 patients taking into account 12 factors, like sex, age and other clinical details. It is important because most recurring colorectal cancer patients die from a recurrent, metastatic disease and not from the original tumour and because the disease returns in about 20% of the cases after initial surgery. Predicting if someone is more likely to relapse after treatment can be vital in choosing the best possible therapy and extend lives. (More on colorectal cancer metastasis)
How does this tool work? Researchers have built a risk model for recurrence from data of thousands of recurring and non-recurring colorectal patients. Based on this, they created a web-based individualised prediction tool that weighs factors that contribute to recurrence. (Find out more)
Will this tool work for patients worldwide? Unfortunately not yet - the dataset used for this analysis came from the US population, and might prove difficult to translate to the specifics of other countries like China or India, due to many factors like lifestyle, diet and others. Therefore, there is a need for consortiums from different countries to collaborate and come together to share data so that everyone can benefit. (Test the online prediction tool)
AI beats doctors in breast cancer diagnosis
AI can now pinpoint breast cancer more accurately than humans.
UK researchers teamed up with Google Health for a recent study based on an algorithm spotting abnormalities on X-ray images.
The researchers used scans of 29,000 US and UK women to see if AI could eliminate more false positive and false negative diagnoses than clinicians.
Re-analysing diagnoses by US doctors, the system managed to reduce false positives by 5.7% and false negatives by 9.4%, i.e. it identified nearly 10% more cancer cases that humans missed.
Deep Learning is likely to gain traction in the diagnosis of lung cancer too - a recent study revealed the AI was able to spot malignant lung nodules on low-dose chest computed tomography (LDCT) scans before radiologists.