MICCAI 2022 Daily – Monday

In traditional supervised learning, we want a model to learn from an image to fit a label and assume the label is correct. Training usually takes place under an ideal scenario using a balanced distribution of images per class. In deep learning for medical image analysis , when large-scale manually-annotated datasets are unavailable, natural language processing tools are sometimes used to annotate datasets with labels extracted from radiologists’ reports. Using this automated process without verification can lead to unreliable or noisy labels. Also, images often require multiple labels and have class imbalances due to the variable prevalence of diseases. “ In medical imaging, typically, most patients are fine, but some have a disease. If you train your model directly on this imbalanced data set, you’ll have a model that always predicts patients to be okay and fails to predict patients having diseases, ” Fengbei explains. “ We use a deep learning method to train a DenseNet model on the given images and propose a new regularization method that, unlike other methods that try to select noisy samples, regularizes the training 20 DAILY MICCAI Monday Poster Presentation NVUM: Non-Volatile Unbiased Memory for Robust Medical Image Classification Fengbei Liu is a PhD student at the University of Adelaide in Australia, under the supervision of Prof. Gustavo Carneiro. His work explores learning with noisy multi- label imbalanced data. He speaks to us ahead of his poster presentation this morning.