Computer Vision News - February 2017

The task of finding automated methods of liver segmentation is quite solved, but lesions are really challenging. Lesions have different shapes and are visible with contrast enhancement, which sometimes vanishes. If we do low contrast, the contract can also be a different kind of contrast. After treatment the lesion stops embolization. There are a lot of manifestations of lesions in the liver. This makes it really challenging especially with methods based on deep learning, which need a lot of training data. Unfortunately, there is not a lot of training data around. This challenge seeks to make it possible for other researchers to work with more training data and hopefully get better results for the lesion segmentation problem. Are there unrelated lesions that look like cancer and might confuse the researcher? Yes, there are structures called cysts . A cyst is a compartment filled with water. Sometimes if contrast is not good enough, it’s hard even for experienced radiologists to detect it. Normally, previous scans of the patient are checked to see if something has already been noticed in this area. It is quite challenging. What will be the practical application of these results? It would be really nice to retrospectively analyze data that has been acquired already. Let’s say every hospital has a HIS (Hospital Information System) or PACS (Picture Archiving and Communication System) where they store the medical data. With this, you can really look back in time and analyze scans that have been acquired already and assess the tumor load. If you then have some relevant clinical data of that patient, you can regress the patient survival from the tumor load. Does this also mean doctors will be able to treat patients faster because they have better understanding of the data? Yes - This could also be the case. The head radiologist can get the full tumor load or tumor burden of the patient. This could be a much better predictor for patient survival than just measuring the diameter of the three largest lesions which is the clinical standard so far. How much data will there be in the challenge? There will be 100 training data and 50 test cases. We will write together and organize the challenge with the best methods in a joint publication where people who are interested can find a summary of the challenge. Once the results are known, we will be in touch with Computer Vision News to share them with your readers. “This challenge seeks to make it possible for researchers to work with more training data and get better results ” Challenge Computer Vision News Challenge 31 “Full volumetric assessment of these lesions will have much better results in describing the state of the patient”