2024 Volume 17 Pages 33-39
Images of mouse brain slices, obtained under slightly different experimental conditions, are available in 84 datasets in the NeuroGT database (https://ssbd.riken.jp/neurogt/). Our goal was to obtain semantic segmentation results for eight brain anatomical regions. However, out of 84 datasets, only one dataset had true labels that could be used to train a convolutional neural network (CNN), and it was incomplete (131 out of 162 images). A segmentation model trained with the labeled images was less accurate on other images obtained under different experimental conditions because of differences of the image properties. We therefore tried Unsupervised Domain Adaptation (UDA), wherein the parameters of the CNN trained on the labeled images (source) were transferred to the unlabeled images (target). We used the positional information of the sample slices associated with each image to propose a novel loss function that approximated the class occurrence probabilities of segmentation results obtained from source and target images of brain samples at similar sliced locations, and we introduced it into the UDA. The proposed UDA method achieved an mIoU of 78.34%, which was 8% more accurate than the previous UDA methods such as Contrastive Learning and Self-Training (CLST) and Maximum Classifier Discrepancy (MCD). We demonstrated experimentally that the proposed method was useful for segmenting biomedical images with a small amount of incomplete training data.
