MICCAI 2022 Daily – Wednesday

In terms of the next steps, Meirui says there is plenty of space for further improvements. The current method can make the model aware of imbalanced class distributions, but there is still some way to go to match the performance of supervised learning with full labels on all clients . “ Our dynamic bank learning currently performs in a semi- and self- supervised way, ” he tells us. “ In the future, we can further explore how to do the dynamic bank construction incorporating information from other clients to solve the potential limitations for handling a more severe class imbalanced issue. Some local clients may not cover all possible disease categories, which means certain diseases are absent for some clients . That ’ s a significant problem in the real world, and we need to solve it as soon as possible because it means some hospitals can ’ t cover all patients across all disease categories. ” A small hospital may not have enough patients to train an AI model on its own adequately. However, when many different hospitals club together with federated learning , particularly in this semi-supervised setting, it means more data , more reliable models , and more confident predictions . Ultimately, clinicians can put more trust in diagnosis. 12 DAILY MICCAI Wednesday Poster Presentation