Google's DeepMind To Apply AI In Head and Neck Cancer Treatments (thestack.com) 17
An anonymous reader quotes a report from The Stack: Google's DeepMind team has partnered with British hospital doctors on an oral cancer program hoping to cut planning times for radiotherapy treatments. After recently announcing a partnership with London's Moorfields Eye Hospital to use its machine learning technologies to speed up the diagnoses of eye conditions, DeepMind has confirmed a new initiative at the University College London Hospitals (UCLH) NHS Foundation Trust. According to Google's artificial intelligence unit, cancer treatments including radiotherapy involve complicated design and planning, especially when they involve the head and neck. Treatments need to obliterate cancerous cells while avoiding any healthy surrounding cells, nerves, and organs. UCLH plans to work with DeepMind to explore whether machine learning can reduce planning time for these treatments, particularly for the image segmentation process which involves clinicians taking CT and MRI scans to build a detailed map of the areas to be treated. The report adds: "DeepMind algorithms will be set to work on an anonymized collection of 700 radiology scans from former oral cancer patients, learning from the historical data in order to draw its own conclusions without human support."
Re: (Score:3)
Re: Hold Up! (Score:1)
Whats the difference between God and a doctor? God knows that he's not a doctor.
AI? (Score:1)
Re: (Score:2)
Just the creepy ones.
Thoughts from therapy imaging (Score:3, Interesting)
Radiation therapy physicist and image segmentation researcher here.
Generally speaking, there has been no revolutionary advance in segmentation algorithm design in the past 5-10 years. Most modern algorithms combine atlas-based segmentation (ABS), with either shape modeling and/or machine learning. By itself, machine learning alone is generally agreed to be less accurate than ABS alone. The data set described in the article (700 patients) is certainly adequate for ABS and shape modeling. It may not be adequate for a pure learning algorithm, but I reserve final judgement on that.
There are a few strange things in the article. (1) A standard head and neck case probably does not require four hours to create a manual segmentation. An hour is closer to correct. Therefore, they have already achieved their goal. (2) An important reason why a complex head and neck case takes longer is to define the therapy target. It is unlikely that simply adding processing power will make this easier. (3) I didn't see any description of what algorithm(s) they will evaluate, nor who will be in charge of algorithm development.