Development and Evaluation of the First Pediatric Deep-Learning Auto-Contouring Models for Cranio-Spinal Irradiation (CSI)

Abstract

Cranio-spinal irradiation (CSI) using advanced modality (IMRT, VMAT, protons) requires contouring of the thecal sac, a tedious and time-consuming task. Additionally, sparing of the vertebrae to improve growth and bone marrow reserve necessitates vertebrae contours. The COG and other investigators have great interest in studying growth effects from vertebrae-sparing technique; this effort requires reporting of dose to the thecal sac and the vertebrae. The presence of bone-soft tissue interface and time required for manual contours make this an ideal use of deep-learning auto-contouring.Two convolutional neural network (CNN) models were trained and validated, one with 65 patients to auto contour thecal sac, kidneys and lungs, and the other with 55 to draw vertebrae. The training set consisted of mostly pediatric and young adult patients. The models were tested on ten new patients at the age of 5, 6, 8, 9, 11, 13, 16, 20, 31 and 47. Quantitatively, auto contours were evaluated against clinical manual contours using dice similarity coefficient (DSC) and mean distance to agreement (MDA). A qualitative scoring system was used by a dosimetrist experienced in CSI contouring: 0 (accepted upright), 1 (accepted, minor editing in 1 min), 2 (accepted, moderate editing in 3 min), 3 (accepted, significant editing with efficiency gain) and 4 (rejected, gross error or no efficiency gain). The time for editing each auto contour for each patient was recorded and compared to the time required for manual contouring measured using time/slice and the number of slices for thecal sac, vertebrae and each kidney, and as whole organ for each lung.As shown in Table 1, on average, the DSC was above 0.90 and MDA below 1.2 mm, both well within the tolerance recommended by AAPM TG 132. The mean score was below 1.0 for all volumes. Editing thecal sac and vertebrae auto contours only required 1 min or less for all patients. The models reduced contouring time by 84 and 40 min for thecal sac and vertebrae, respectively.The deep-learning models provided high-quality auto contours, directly acceptable for lungs and requiring only minor editing in less than one min for thecal sac, vertebrae and kidneys. Given the labor-intensive nature of manually contouring thecal sac and vertebrae, the high accuracy of the auto contours resulted in substantial reduction of contouring time and thus faster turnaround for CSI treatment. The models can provide a uniform starting point for CSI contouring in our field, which is particularly beneficial to clinics with less experiences.