Room: Exhibit Hall | Forum 9
Purpose: When apply deep learning to deal with three-dimensional medical image, owing to insufficient training samples and complicated anatomical environments in volumetric images , the problems of degradation and gradient disappearance of 3D convolutional neural network optimization appear. Here, we develop a fused model Dense V-Network to solve the problems above and achieve accurate delineation of pelvic organs at risk (OAR).
Methods: We combine two typical network models of Dense Net and V-Net to set up our algorithm Dense V-Network. For training model, we adopt 100 plan computed tomography (CT) images of patients with cervical cancer, 80 of which are randomly selected as the training sets to adjust the parameters of the automatic segmentation model, and the remaining 20 are used as test sets to evaluate the performance of the convolutional neural network model. Whatâ€™s more, we take three representative parameters to quantitatively evaluate the segmentation effect.
Results: The clinical results showed that the Dice similarity coefficient values of the bladder, small intestine, rectum, femoral head and spinal cord were all above 0.87 (average is 0.9); Jaccard distance of these were within 2.3 (average is 0.18). Except for the small intestine, the Hausdorrf distance of other organs were less than 9.0 mm (average is 6.1 mm). We also compared our approaches with results with those of Atlas and literature: Dense V-Network showed performance and faster speed.
Conclusion: The results above suggest that Dense V-Network algorithm can be used to automatically segment the OARs accurately and efficiently. In addition, it shortens the patientsâ€™ waiting time and accelerates radiotherapy workflow. As the algorithm structure is optimized and the training data increases, the expected results will be further improved.