Automated Clinical Target Volume Delineation for Non-Small Cell Lung Cancer Patients Using Deep 3D Networks

Abstract

Patient-specific clinical target volume (CTV) delineation is the most difficult step in defining the planning target volume from the gross tumor volume. Delineating patient specific CTVs is a labor-intensive process, prone to human error, with poor inter-observer consistency. Automated volumetric expansions have been used to reduce these concerns, but these ignore patient specific anatomic and clinical variables, as well as the clinical experience of the radiation oncologist. We developed an automated CTV delineation system for lung cancer, which was trained to learn the physician’s CTV contour pattern using a deep 3D neural network. Our network took both patients’ CT images and their corresponding internal target volume (ITV) as input, and output CTV delineation. ITVs were contoured by physicians after studying the internal motion of the target from the 4D CT scans. In our study, the network was trained from twenty-five NSCLC patients in right upper lung lobe (RUL). Then, the trained network was tested on four additional datasets from RUL NSCLC patients. Each CTV delineation output by the trained network was compared with the corresponding CTV contoured by physicians. One challenge to train a 3D network for CTV delineation is the high imbalance between the CTV volume and the background volume. Instead of using hyperparameters to balance as existing methods did, we directly optimize a loss function robust to high imbalance. This improves the stability of the model, especially when the balance dynamically changes in images. Three out of four CTV contours that were automatically delineated by our trained model agree very well with the physician’s CTV contour. Only one has visible discrepancies near the apex of the upper lobe, but still looks reasonable. As shown in Table 1, the performance of the trained neural network was also quantitatively evaluated using precision and recall between the predicted CTVs and the ground truth CTVs. The average precision is 0.902, and recall is 0.871. The recall and precision are the fraction of true volume among the total ground-truth and the fraction of true volume among the predicted volume, respectively. The automated CTV delineation was utilized for RUL NSCLC patients using the deep learning method solely based on CT images with manually contoured ITV. On the limited testing samples, the model contoured CTVs reasonably well. This approach would potentially combine the ease of automated volumetric expansions while incorporating the patient specific anatomic and clinical variables along with the clinical expertise of the radiation oncologist.Tabled 1Abstract 2782; Table; A quantitative evaluation of our automated CTV delineation system on four testing patients.Test #1Test #2Test #3Test #4Recall0.9420.7270.9090.907Precision0.8150.8890.9550.950 Open table in a new tab