Computer Vision News - November 2022

45 Medical Computer Vision of CT images from low-dose data. In this case the goal is to reduce the patient exposure to ionized radiation. While deep CNNs have demonstrated unprecedented performance in many image restoration tasks. However, their operation is inherently local, making them less suitable for handling CT reconstruction tasks, which suffer from global artifacts. While existing works circumvent this incompatibility using deeper networks with very large receptive fields, Bassel addressed this issue from the data perspective by proposing a new locality-preserving representation for the CNN’s input data. Other oral presentation topics included incremental 3D medical image segmentation using Multi-Scale Multi- Task distillation and predictive model to estimatewithdrawal time in colonoscopies. In addition to the talks there were 19 poster presentations on a variety of MCV topics which intrigued intense discussions. annotated with only 5% of the cells in each tile. To tackle both cell detection and classification within a single network, she modeled these as a multi-class segmentation task, and trained the network with a combination of partial cross-entropy and energy-driven losses . It is encouraging to see that two out of the three first authors for the best papers are female. Additional scientific oral presentations presented addressed the challenges of medical image reconstruction. On MRI reconstruction, the goal is the reduce overall acquisition time by undersampling the data frequencies. The novelty in the work presented by Aniket Pramanik from the university of Iowa resides in the combination of segmentation and reconstruction tasks to improve the model performance rather than focusing on the reconstruction itself and perform the segmentation as a subsequent task. Another work by Bassel Hamoud from the Technion focused on reconstruction T OF CCV