A Fully Automatic Framework to Localize Esophageal Tumor for Radiation Therapy

Automatic localization of esophageal tumors is an important part of target volume planning in radiotherapy. Currently, the main localization method is manual localization. Traditional manual positioning is time-consuming and inaccurate for the following reasons. First of all, esophageal neoplasms are irregular in shape. The second, the tumor image was insufficiently contrasted with the surrounding tissue. Also, the tumor area is highly heterogeneous. To solve these problems, this paper proposes an automatic positioning framework combining single point multi-box detector (SSD) with the optimized VGG16 deep learning network. The optimized algorithm network has achieved good results in our esophageal tumor localization experiment. The experimental data consists of 96 esophageal VMAT plans and training set consists of 60 patients, the remaining 36 patient data sets were used as the test set. We trained with 5000 slices and tested with 1000. The experiment result showed the tumor areas of 820 CT slices were effectively located, and the accuracy rate of intersection greater than and (IoU)[6] value was 82%. These promising results suggest that the target area of esophageal tumor can be well located in our optimized framework, which can improve the efficiency and quality of plan making of esophageal tumor radiotherapy.