MICCAI 2022 Daily – Wednesday

“ Some works try to apply traditional semi-supervised methods into a federated learning scenario, ” Meirui tells us. “ Those methods use techniques like consistency-based learning or pseudo labels , but the class distribution is typically imbalanced because of different demographics and disease incidence rates in medical applications. We want to find a way to solve both the semi-supervised learning and the class imbalance issue . That ’ s the main challenge here. ” The team designed a new dynamic bank learning method , which iteratively collects the unlabeled samples and assigns each with a pseudo label. By doing this, it can progressively accumulate more samples in the bank. Those samples are considered to be confident samples. “ Using these confident samples, we designed a sub-bank classification task, which helps our model to learn the class discriminative knowledge, ” Meirui explains. “ With this scheme, we ’ re able to make the model aware of different imbalanced distributions so that it performs better on the clients without any annotations. ” Supervisor Qi Dou told us that she is particularly fond of this work because it shows the possibility of effective federated learning under extremely minimum annotation cost, which will lower the barrier for collaboratively conducting large-scale privacy-preserving study among medical institutions in practice. And why does Meirui think Qi is so proud of his work? “ I think it ’ s because we ’ ve chosen a practical problem faced in real- world scenarios, ” he responds. “ By using this cherished unlabeled data, future works and models will benefit from a large distribution of data and perform better for certain medical imaging tasks. ” 11 DAILY MICCAI Wednesday Meirui Jiang