לפצח את אתגר האבחון הרפואי: פריצת דרך ב-AI

Thank you all. It's a pleasure to be here. Quite an honor, actually. I'll be presenting the research I did in my final year, mentored by two professors. One is right here in the other room. She's chairing the AI and Medical Bella Spector and Dr. Ilan Shimshoni. Okay, cool. All right. So medical imaging and segmentation. I don't really have to go into the real world application of it, but I will mention that it preserves challenges and advantages in this kind of frame. They tend to be not noisy, very predictable, very straightforward imaging. You tend to need less circumstances samples to train a proper model. But labeling is quite an issue. You require a medical professional to label the images and you require them to be pixel perfect. Eventually, for example, if we look at that image that's presented behind me, best I could do is like, this is a brain. Medical image can create 18 or label 18 different structures in the brain with how efficient and how efficient this can be and how each label is expensive to create. We very easily landed a solution that active learning would be very useful in this kind of case. But active learning presumes an active learner that initially would know to examine and to decide which labels are worth pursuing later and which are not. This is what we call the cold start initiation or how we first achieve this initial model and what samples we train it on. Common example would be to train it on random data points. If we look at the left, we'd see that it could be risky. We could over represent some of the data, not represent some of the rest of the data. And what we would like is something closer to this, where we have our whole data is represented in a fairly consistent way. So how do we do this or what we suggest to do? So we propose a pipeline for the initial cold start problem. We start obviously with our unlabeled data and we use a foundation model, what we might call a foundation model, but it's a large model trained on as much data as we can, millions, presumably on different modalities. This way we can use it for all kinds of different tasks. So it would be trained on MRI, X rays, CTs, anything like this. And we use this foundation model to extract embeddings out of unlabeled data. Following this, we reduce the dimensions. Different methods exist. We will present one of them while we followed several cluster and then pick our samples out of them. Part of what we this kind of approach has been done before and what we provide is actually how many clusters to select and how to select the samples from within each cluster. Also it has a nice visual effect to understand. So I'll show you right now. So we would start here with our embedding unlabeled data, post embedding, post emancipation reduction. And we would try several different amounts of K. For our K means, we'd use silhouette score, which is for those unfamiliar, the kind of measure that measures how well the clusters are within each other and separated from other clusters. So we'll see, we can try several. And we select our highest. Our highest count is what we decide to choose and continue with. Once we do this, we select the medioids of each cluster. The medoid of each cluster is the actual data point which minimizes its dist distance to all other center to all other points in the cluster. And we'd presumably have some extra budget. In our case, we wanted overall 15 samples. So right here, since we chose 10 clusters, we allocate each cluster its furthest point, trying to maximize the difference within the cluster relative to the cluster size. So larger clusters get more represented. We get something like this. We see how well we can follow our potential data. Right. Before I go into the results of this, I wanted to show one more thing that we researched and one more step is how we can leverage this embedding space. Obviously active learning is being widely used. This is why I brought it up in image segmentation for medical uses. But we fall into some kind of issue, especially if we would select based on entropy, which is a common approach. Especially if we want to select only a few samples. For example, here, if I want to add five samples, we see them in blue right here. This happens. I select three very, very similar samples and I risk redundancy. I would like to have better representation. What we did is leverage already embedded space and combine the scores of the entropy, normalize them with its distance. And you can see right here. Even with five samples, I managed to drop one redundant sample at the bottom and add one more sample at the top left. Better representing and still very high entropy. And let's go into what we actually researched. So we did our research on three different data data sets with different modalities, which was important to us to try both MRI and X rays because we could use the same foundation model. We could prove it's efficient on both types of modalities and different kind of tasks. All the technical stuff at the right, I will skip. Okay, here we have our results. I know it's a lot to look at, but I recommend focusing on the blue boxes at the left compared to the relevant competitions. You can say which is both a random baseline and clustering without reducing the dimensions. Clustering without reducing the dimensions is an approach you could find but it tends to be although it could work well it tends to be very noisy and can presumably weaken our results so we wanted to show ours better and on the purple we see the combined scored active learning step compared to entropy and maximizing the distance. All right mostly what we found is clustering the cold start initializations makes a big difference, works better than competitive baselines and makes the biggest difference at low budgets when very few structures when very few samples are selected then then it makes a bigger difference and tends to plateau towards higher percentages of data. As I mentioned previously you don't need that much data to train a good model on this kind of tasks. That's about it. Thanks to Bella Spector like I mentioned and Ilan Shimshoni were both my professors following me in this research and if there's any questions obviously.