• Light
  • Dark
  • Auto
Select Page

In her last episode, Prof. Andrea Rockall, from the Imperial College London, was joined at the London, England studio by Theo Barfoot, an MRI physicist at the Royal Marsden London, and Ben Glocker, from the Department of Computing at the Imperial College London, both of whom have worked with Rockall for many years.

Even before the episode began, there was a frenzy of activity in the chat. The audience was asked to participate in a poll right at the beginning of the show. They were asked if they had experience in whole body MRI (WB MRI); 109 of the viewers participated, 6% of whom had a lot of experience, 12% moderate, 32% low and 50% had no experience whatsoever.

Andrea Rockall explained that WB MRIs can be very difficult for radiologists to read, which made the validation in this use-case a challenge; nonetheless, it is a good imaging technique for any metastatic issue that could potentially affect the whole body. She then presented the clinical issue of patients who are suspected of having colon cancer after a colonoscopy and are then transferred to the radiology department for staging, which results in numerous examinations. Traditionally, this can be a very time consuming and tiresome process for both the patient and the clinician; WB MRI would save time. However, as the complexity of reading whole body MRI scans is seen as a burden for the radiologist, Rockall proposed once again creating an algorithm which can support the clinician in identifying malignant lesions on whole body MRIs. If successfully implemented, this could turn such an examination into a “one-stop-shop”.

When asked about the issue, Theo Barfoot said that the key sequences to map cancer, in his opinion, are DWI, T1- or T2 mapping, which provide very rich multi-parametric data sets that can be useful for machine learning.

Rockall presented two recent articles underlining the effectiveness of whole body MRI in staging colorectal and lung cancer. Ben Glocker mentioned that without AI, whole body MRI might not be a successful tool in the clinical practice. Barfoot added that radiologists are required to create a training set and label the data in order for the algorithm to work.

Then, Rockall led the audience through the workflow requirements when setting up the database. Her project was multi-centre with 15 different centres and thus, multiple vendors, in order for a diverse data set to be established. She showed the output of machine learning tools that segmented organs in whole body MRIs. The audience found out which organs were easily detectable for the machine and which ones were a challenge, thus requiring larger training sets. Rockall explained why data harmonisation is crucial in machine learning and why the performance of an algorithm drops when the input data changes. Glocker then explained how classification forests work.

Next, Rockall presented the segmentation process and the associated challenges. Theo Barfoot and Ben Glocker added that, for example, sometimes in a training set in WB MRI, heads have to be added if the head was not included in the image, so the algorithm is able to understand what it is looking at. Rockall explained why the test set performs best if it is trained on a substantial data set and that it is important to calculate statistically how many patients are needed for the set to ensure this. The entire data set needs to be looked at, stratified and balanced before randomly dividing it. Glocker went on to explain how to best go about dividing the data into training-, validation- and test-sets.

Andrea Rockall depicted the general circle of machine learning before she presented the “happy end” of her machine learning project and some of the challenges in clinical validation.

The final Theo’s Technical Top Tips was presented live at the studio and covered model validation (testing) in the lab and in a clinical environment.

Glocker then talked about the deployment of algorithms and implementing them in the actual clinical practice. Rockall underlined that one should always remember that there is a patient at the end of every machine learning tool and that it is crucial for it to be safe prior to implementation.

As always, these dedicated speakers spent extra time at the end of the episode to make sure that all chat questions from the viewers had been answered.

All episodes are now available on demand for purchase on

Read more on Episode 1, “The one where we classify breast cancer with MRI” here.
Read more on Episode 2, “The one where we tackle fibrosis” here.
Read more on Episode 3, “The one with the sixpack” here.
Read more on Episode 4, “The one that might be renal cancer” here.
Read more on Episode 5, “The one where we find the connection” here.

Want to deepen your AI knowledge? Check out the reading recommendations from our experts here.

We sincerely hope you have enjoyed this season of ESR Connect Weekly, “Use Cases in AI – Reasons to do AI with Friends“. We would like to thank all of the speakers, hosts and technical team who devoted their time and energy, putting on such high-quality episodes each and every week across different countries. And, of course, we would like to thank the viewers for participating in the chat and engaging in the episodes!

Latest posts