Computer Vision News - January 2024

39 Computer Vision News Datasets like Cityscapes collect real data with sensors attached to cameras to train the algorithms. However, collecting real data for medical scenarios is challenging due to the high annotation cost and logistical difficulties in obtaining data during surgeries. There is not a large volume of data available that can be used to make realistic scenes. SARAMIS seeks to overcome these hurdles by providing the first set of environments and data that can be used to generate a wealth of synthetic data, enabling the training of computer vision algorithms for medical applications. It comprises 3D meshes, textures, and tetrahedral volumes covering various anatomical targets. It can create environments for training reinforcement learning (RL) agents to perform specific actions or generate realistic images with paired labels, such as segmentation maps, depth maps, and optical flow. “SARAMIS was acquired from a relatively large set of open-source CT scans, which were previously available,” Nina tells us. “The dataset is made up of over 2,500 individual patient cases where all these anatomical targets were segmented, reviewed, and then converted into assets. We have 2,500 now, but realistically, if another large set of CT scans was openly available, we could apply the algorithms and methods to more scans to obtain an even larger set of these assets.” Currently, the dataset encompasses 106 anatomical targets, including two previously unlabeled in the literature. The team used an open-source SARAMIS