3D FULLY CONVOLUTIONAL NETWORK FOR THORAX MULTI-ORGANS SEMANTIC SEGMENTATION

Automatically delineating Organs-at-Risks (OARs) on computed tomography (CT) has the benefit of both reducing the time and improving the quality of radiotherapy (RT) planning. A 3D convolutional deep learning framework for multi-organs segmentation is proposed in this work; moreover, for the small volume OARs, a robust 3D squeeze-and-excitation (SE) feature extraction mechanism and a new Dice loss function are incorporated in the traditional 3D U-Net. We collected 60 thorax CT images set with annotations and expanded to 260 patients by the augmented method of randomly rotating +/- 6 degrees with a 1/3 probability and adding Gaussian noise. The objective is to segment five important organs: esophagus, spinal cord, heart, and bilateral lungs. Compared with 3D U-Net, 3D-2D U-Net proposed in our work increases the Dice similarity coefficient by 5% on average for the heart and bilateral lungs, and 3D Small Volume U-Net can further increase the Dice similarity coefficient to above 80% for the spinal cord. The experiment results demonstrate that the proposed model can improve the delineation accuracy of OARs from CT images.