Computer Vision News - January 2023

37 Skylar Stolte A big struggle for coders is repositories that aren’t documented well. MONAI is really well documented, and it has all of these neural networks to work with and pre- and post-processing functions that work well on a dictionary of all of your inputs and can handle any problem you have. The libraries available under this framework are also GPU-accelerated frameworks , so we have this faster and more efficient software that allows us to focus on solving future problems rather than struggling with basic challenges. ” When CVN spoke to the guys at NVIDIA about MONAI, they promised something almost revolutionary, and it seems they kept their word! Skylar says that because MONAI has been so useful for this work , the lab wants to give back by contributing to progressing the tools available for other researchers worldwide. “ We presented our loss term in our MICCAI oral talk this year, and we’ve been talking about the potential for implementing it in MONAI, ” she reveals. “ This would help researchers because although there are many efficient losses in MONAI, none are calibrated losses. Calibration is a field of deep learning that’s not talked about enough. ” challenging to translate the performance to the data from the other location. “ It’s fundamental in getting deep learning in medical practice that you have to be more translatable to different locations and patient populations , ” Skylar asserts. “ That is where the idea of the calibration approach comes in. Calibration allows you to get an output in addition to your classification that measures your confidence in your prediction. When we have this additional output that tells us how likely we are to be right or wrong, we can use it to take an output where we think that our model isn’t very certain and have manual segmenters help us improve the labeling, feeding that back into our model to train it to handle these harder cases better. ” The lab has been using MONAI (Medical Open Network for AI) software, which should be familiar to our readers, and was developed from a project originally started by NVIDIA and King’s College London . “ MONAI has been so helpful in my research because it comes with many frameworks already implemented in MONAI, ” Skylar tells us. “ You can take the U-Net from MONAI and use that in your code instead of a random PyTorch repository. It allows you to take these different frameworks, easily change the parameters, and get your code running without too much effort .