Rivals in the ongoing big data/artificial intelligence (AI) challenge that will take place during the course of Journées Francophones de Radiologie Diagnostique et Interventionnelle (JFR 2018) are eager for the next round of images, scheduled for distribution on opening day, to be held in Paris on 12 October.
It's hoped the challenge will strengthen collaboration between players across the whole AI ecosystem in France.Le Grand Challenge, or Data Challenge JFR, organized by the Societé Française de Radiologie (SFR), hinges on two training rounds almost a month apart, and a final validation round on Sunday at JFR. The goal is to train algorithms to tackle five separate clinical topics: identifying a meniscus tear on MRI, segmenting renal cortex on CT, identifying hepatic lesions on ultrasound, detecting breast lesions on MRI, and thyroid cartilage lesions on CT.
"It's likely there will be a few sleepless nights for the team members as they hone the programs," said chief organizer Prof. Nathalie Lassau, professor of radiology at the Gustave Roussy Institute in Paris. "The competition is 'contre le montre' (against the clock) and so a dedicated workspace at JFR is open 24 hours a day between Friday and the Sunday deadline at 2 p.m."
Supported by computer engineers from the CentraleSupélec academy, Théo Estienne and Philippe de Vomécourt, Lassau and her team of six organizers invited hospitals to recruit patients between 15 May and 15 September. A total of 45 both public and private hospitals participated, recruiting 3,500 patients and contributing their image data -- each clinical question comprising a minimum of 600 images.
Prof. Nathalie Lassau from Gustave Roussy Institute.Dataset availability varied. For example, there were around 1,500 knee MRI exams compared with 600 liver ultrasound exams. The organizers provided the first third of the dataset for each challenge, along with clinical information, to each competing team on 15 September so they could start training their algorithms.
When JFR 2018 begins, teams will access the second third of the dataset, complete with patient diagnosis, for each challenge in order to fine tune their algorithms over two days before the final task on Sunday at 2 p.m.: In this final round, teams must validate their algorithms on the last set of 200 images.
This time their programs will be blinded to patient information, and the teams will be judged on the efficiency of their algorithm. The judges, including Profs. Anne Cotton and Marc Zins, will decide which algorithms are the best detectives for identification of normal and abnormal findings, and in each category, the winners -- to be announced on Monday -- will receive 3,000 euros and have their work published in Diagnostic and Interventional Radiology.
So who are the teams?
Initially the challenge aimed to create an inventory of current AI actors, and more importantly, promote and strengthen collaboration among these stakeholders. Each team must comprise at least one radiologist, an engineering student, and a company or research lab. They include researchers from institutes like INSERM and Paris Sud, computer scientists from academies such as CentraleSupélec, and industrial partners -- both large, established companies such as Canon, GE, IBM, and Philips, and also smaller start-ups, such as Owkin and Illuin. While most of the teams hail from France, there are a couple from other countries; one from Switzerland and one from the U.S.
Not all teams are addressing the five topics. Some are concentrating on one, or several questions, depending on their focus of interest and associated expertise, according to Lassau. Similarly, the teams, constituting 180 participants in total, range in size from small groups of three to larger teams of 10 to 15 members.
"The Grand Challenge is an exciting prospect. We've never had five challenges in one congress before, and it's a great way to put French AI on the map," Lassau told AuntMinnieEurope.com ahead of the meeting. "We aim to make full use of the dedicated AI space at JFR, which with its three areas, simultaneously serves as an open theater for sessions and pitches, a workspace for the challenge, and booths for start-ups."
JFR's AI sessions begin on Friday afternoon with a summary of AI's importance in 2018 between 2 p.m. and 4 p.m., in tandem with the delivery of the second round of images to rival teams. More two-hour sessions are scheduled over weekend, including AI for idiots at 10 a.m., AI in radiology at 2 p.m. on Saturday, AI at the heart of patient care at 10 a.m., and Ethics and regulation in AI at 2 p.m. on Sunday. The Data Challenge prizes will be awarded at the AI impetus in imaging session on Monday between 10 a.m. and noon.
Are there any hot favorites? Of course, France is famous for its impartiality, but if pushed, the organizers admitted the teams with start-up partners were the ones to watch.
"There are no results as yet, but start-ups offer participants agility, and it's easier to get everyone working around the same table more quickly. I feel sure we are going to see some great results from teams working with smaller companies," de Vomécourt noted.
For more details about the challenge, click here.
