Diaphragm-Net: 3D Full Automatic Segmentation of the Diaphragm Based on Deep Learning Methods for 3D-CT Images

Predicting breathing movement is complex and is regarded as one of the most challenging areas in medical research. The diaphragm is the main muscle of respiration, thin and flat structure, making its automatic segmentation difficult. Manual segmentation is a tedious, time-consuming task, and expert segmentations are prone to intra- and inter-operator variability. The full automatic segmentation of the diaphragm remains difficult and challenging due to its thickness, irregular shape and the similar density with the surrounding organs. In this work, we have developed a new automated deep learning (DL) segmentation framework of the diaphragm. More precisely, we have proposed a new multi-axis annotation refinement workflow using deep transfer learning (Multi-Axis-DTL). This relies on deep transfer learning of the 3D nnU-Net model to assist and refine expert annotations on all the 3D-CT images axes, reducing annotation cost. Our results showed that the nnUNet-3D-enhanced model consistently outperforms the 2D/3D nnU-Net models. This improvement is due to segmentation constraints imposed by additional classes, which provide valuable contextual information to the model. We show that models trained on the produced diaphragm dataset show promising results while training on a limited number of 3D-CT images.