Deep learning based multi organ segmentation in computed tomography images

With the evolving technology, the most significant advancements in the field of health are happened through medical imaging techniques. As a result of complicated and detailed imaging of the within structure of our body with help of medical imaging techniques, data about the state of the organs is received. These obtained images are read, rendered and interpreted by radiologists. Determination of organs and tissues in medical image analysis is the first step of disease diagnosis and cure planning. It is hard and time-consuming to recognize organs through increasing medical images. In this study, a computer-assisted automatic diagnosis system that provides segmentation of more than one organ on computed tomography images of the abdominal region has been implemented to assist radiologists. A fully convolutional neural network, which is a deep learning method of automatic multi-organ segmentation process, is used since Deep Learning models have accomplished high achievements in the segmentation field same as other computer vision fields. In this study, the data set published for the multi-organ segmentation contest (MICCAI 2015 Multi-Atlas Abdomen Labeling Challenge) from Vanderbilt University has been used. The files in this dataset are 3D abdominal tomography images in NIfTI format. A fusion approach is recommended using a two-stage 3D U-Net model that combines different color spaces by acquiring images in HSV color space from these images. Dice similarity coefficient was used to assess the proposed model, and the organ with the highest Dice score as a result of the test procedure was the liver and the organ with the lowest score was the left adrenal gland. Considering the average accuracy score of all organs, it was seen that the automatic segmentation system implemented to help radiologists is successful and promising.