ISBI Daily - Friday

Poster Presentation: Deep Patch-Based Priors under a Fully Convolutional Encoder-Decoder Architecture for Interstitial Lung Disease Segmentation 14 Friday ISBI DAILY Their work explores automatic segmentation of a lung disease called scleroderma, one of a group of diseases called interstitial lung diseases. Guillaume tells us that in radiology, it is difficult to recognise and quantify the disease. When patients are seen across several appointments, tracking the course of the disease through images to see if it has worsened or stabilised – it rarely improves – can be problematic. It is therefore very useful to have an objective evaluation of the extent of the disease. He explains that currently, the aim is not to detect the disease earlier, because to try to detect the disease the patient must have symptoms, and as soon as they have symptoms it is already too late. It is not possible to prospectively find which patients could have the disease, only to look at its evolution. Many people are already working on lung disease detection. There is a publicly available dataset called the interstitial lung disease dataset which contains a lot of different types of disease, but they are spatially annotated. This means they have sliced annotation, but it is not clear if all of the regions in the images have the disease. Generally, this is a problem for medical imaging because annotations are difficult to obtain, or even if there are annotations, they are from one doctor and then a model is learned on one doctor and there is bias in the system. This work is finding a way to combine datasets even if they are annotated in different ways. For example, with a sparse annotation, which means that only the regions that confidently have a disease are annotated. Guillaume has been working on a full annotation dataset, with pixelwise annotations, Guillaume Chassagnon is a PhD student and a radiologist. He speaks to us with Maria Vakalopoulou ahead of their poster session today. They both work with Professor Nikos Paragios at CentraleSupelec in Paris. Guillaume Chassagnon “ The chance to combine own annotations with others which were done in different conditions and are made available online is a major advantage! ”