ISBI Daily - Saturday

18 Saturday ISBI DAILY Ron, Wednesday was my first day at ISBI, ever. Can you tell me what ISBI looked like in 2002, when it started? The meeting actually started in the Washington, D.C. suburbs in Northern Virginia, and it was a very exciting time in medical image analysis. In fact, there was so much demand for a conference that focused on medical image analysis that even the first meeting was successful. I remember meeting for the first time some new colleagues who I continued to see at ISBIs over the years, all around the world. For me, it’s been a wonderful community of people that I’ve seen time and again, and new people that I’ve met. I think one of the reasons that I come back again and again to this meeting is because of the people that I meet. Not only the people that I know from over time, hearing about how their careers have changed and what their thoughts are about the field as it has changed, but also the new young people coming into the field who are bringing excitement and enthusiasm and new ideas. How has the meeting changed or stayed the same over the years? One of the recurrent themes that I’ve experienced over time at meetings such as ISBI is that new techniques come into vogue, everybody pours into the area to see how well it’ll work on their problems and tries tweaking it and adapting it in all the ways that they can think of. This is a recurrent theme that I’ve seen time and again. Can you think of something that you have learned from ISBI in recent years, that you didn’t expect to take home with you? One of the things that I’m seeing more and more at the meeting is the use of a variety of publicly accessible datasets. For example, I attended a talk within the last hour in which somebody developed a new technique and then applied it to four or five different public datasets, showing a common theme in which his team’s method led to improvements across all the different datasets. Another theme that I think transcends all the talks and all the ISBI conferences is the one of validation. I’m frequently wondering, as I listen to a talk, whether the improvements are actually significant or whether they’re results of chance and tweaking of the technique until success is achieved. I particularly say that because, going back to the challenges that I just mentioned, I frequently see that the challenge winners often have very different methods and yet achieve similar performance. I often wonder whether differences are simply due to chance rather than actual algorithmic improvements. What would you advise a young student to do to be assured of a rigorous approach? It’s very important to use datasets that have as little bias as possible. For example, selection bias, where easier cases are chosen. It’s also very important to be sure to never use the Ron Summers