“Deep learning algorithm performs similarly to radiologists in the assessment of prostate volume on MRI” and interview Dr. Erik Thimansson
Hi Erik, who are you and what do you do?
I live with my wife Åsa and three kids (Klara 8, Hjalmar 12 and Albin 16) in Helsingborg, Sweden and work as consultant radiologist in Helsingborg Hospital with gastrointestinal and genitourinary radiology with focus on MRI prostate. I do part time research as PhD candidate at the Institution for translational medicine, department of radiology in Lund University and my thesis will be about AI-applications in MRI Prostate and prostate cancer screening. I have a long term relation with TMC dating back to 2010 when I worked three years in the Barcelona office (body section) and I also had the privilege to work in Sydney office 2018-2019 (emergency section).
Can you give us a little bit of background on your thesis/reseach?
Yes, of course. The paradigm shift in the prostate cancer clinical workup to ´MRI first´ (meaning MRI prostate is performed before biopsies) increases the number of MRI´s and we already today face a lack of prostate specialised radiologists. Also, high quality reporting is extremely important to avoid overdiagnosis of indolent cancers (leading to unnecessary biopsies and over-treatment) at the same time avoiding underdiagnosis of significant cancers. Volumes also increase with organised testing for prostate cancer which was recently recommended by the European Union Council.
Why did you perform this study?
We will have to use our radiologists in a smart way and make sure the radiologist performs complex work tasks. If radiologists use their time for simple, time-consuming tasks there will not be enough radiologists! For the radiologist community to cope with this I think we need help from the artificial intelligence algorithms, and several AI products are already on the market. But we need studies to evaluate and validate the AI-models in true clinical settings, a CE marking/FDA clearence does not mean the AI model works in real life!
One tedious and time demanding task for the radiologist is to calculate the prostate volume on MRI where today´s gold standard method requires three measurements by the radiologist. We asked the question ´can AI assess the volume as good as the radiologist in a multi-center multi-scanner MRI dataset?´.
You just published “Deep learning algorithm performs similarly to radiologists in the assessment of prostate volume on MRI” in the European Radiology Journal
What are your most important results and conclusions from the study?
We show that commercially available AI can assess prostate volume at least as good as the radiologist using today´s gold standard method. Our method agreement is based on both expert radiologist planimetry (delineate prostate contour on several images) and prostate specimen weight and our dataset is diverse and mimics a true clinical context (124 patients, 8 hospitals, 7 scanners, 1,5 &3T). The result is a promising step towards algorithms helping reallocate radiologist resources towards more complex work tasks than manually measuring prostate volumes.
What is your future outlook – will AI replace radiologists?
No, I am pretty sure that is not a probable scenario. I believe there will be a demand for a ´human-in-the-loop´. We radiologists will work together with AI-models, letting AI help us out with less complex work tasks so that we can focus on our core – create value for the referring clinician and the patient.
Let´s take prostate cancer as example! The most important decision will be made at the multidisciplinary meeting: operation, radiation therapy, drug therapy or a combination? AI will probably be involved in several pathways ( MRI prostate, digital histology, individualised risk calculations etc) but the radiologist must take responsibility for the final staging and the medical doctors at the conference (all being humans, urologist, radiologist, oncologist, pathologist) will together land in a final decision!