Automatic bed removal from small animal CT images using deep learning based segmentation

Numerous medical imaging applications necessitate removing the subject’s bed from X-ray computed tomography (CT) images. Here we propose and explore a novel automatic approach for bed removal in small animal CT scans. The solution uses a 2D neural network for semantic segmentation of the rats. To the best of our knowledge, this is the first time a CNN approach has been used for bed segmentation/removal. The entire dataset used for this study included whole-body CT scans of thirteen rats taken in two different sites by the same scanner model but with different beds designs. After finding the most suitable network architecture, loss function, and training duration (epochs), we performed a 5-fold cross-validation on the remaining dataset acquired at the first site. The results demonstrated high compatibility between the network output and the ground truth labels, with an intersection over union (IoU) value of 0.943 and a Dice similarity coefficient (DSC) value of 0.969. However, when training on datasets acquired on the first site and testing on the dataset taken at the second site, the suggested solution didn’t work as well, most likely due to the differences between the datasets.