Fusion Segmentation of Head Medical Image with Partially Annotated Data

摘要

In this paper, there are 2 head computed tomography (CT) scan datasets: one with vision related organs pixel annotations and the other with audi-tory related organs pixel annotations. We aim to train a single neural network for vision related organs and auditory related organs segmentation at the same time with these 2 partially annotated datasets. An idea generating from co-operative training method will be applied to complete the lacking annotations of each dataset. However, it is not a proper way to treat the predicted annotations as the real annota-tions from professional doctors. To address this error, a semi-supervised method is chosen. Compared to the baseline method, our training pipeline is able to complete 2 segmentation tasks within only one model, and we have proved that it outper-forms the baseline method. To some extent, using partially annotated medical image datasets can help to solve the problem that the scarce source of profes-sionally annotated medical image data. What's more, the proposed method will achieve better performance.