Combining Prior Day Contours to Enhance Automated Prostate Segmentation

Abstract

To improve the accuracy of automatically generated prostate, rectum and bladder contours, required for daily online adaptive therapy. This is achieved by using contours from image guidance (IGRT) scans of prior treatment days to enhance the automated contour generation of the current scan via the proposed algorithm. Six IGRT prostate patients with daily kilovoltage (kV) cone beam CT (CBCT) had their original plan CT and 9 CBCTs contoured by the same physician. Three types of automated contours were produced for analysis: (1) STANDARD: by deformably registering the plan CT to each CBCT, and then using the resulting deformation field to morph the plan contours to match the CBCT anatomy; (2) PRDAY: the contour set drawn by the physician on the previous day CBCT, is similarly deformed to match CBCT anatomy; and (3) STAPLE: the contours drawn by the physician, on each prior CBCT, are deformed to match the CBCT anatomy to produce multiple contour sets. These sets are combined using an algorithm, STAPLE, into one optimal set (e.g., day 3 CBCT combines contours produced using the plan plus day 1 and 2 CBCTs). STAPLE computes a probabilistic estimate of the true contour from this collection of contours based on maximizing the sensitivity and specificity of each organ's contour. ABAS (CMS Inc.) was used for this purpose. It utilizes an intra-patient deformable registration algorithm with a normalized sum of squared differences block matching phase followed by dense local correlation coefficient based registration, implemented on a Graphics Processor Unit (GPU). Compared to STANDARD, and PRDAY, STAPLE improved the average Dice's Coefficient (DC) between the original physician drawn CBCT contours with DC of; bladder: 81 ± 13%, 91 ± 6%, and 92 ± 6%; prostate: 75 ± 8%, 82 ± 5%, 84 ± 5%; and rectum: 79 ± 6%, 81 ± 6%, 85 ± 4%, respectively. The STAPLE results are within intra-observer consistency, determined here by the physician blindly recontouring a subset of CBCTs, e.g., prostate 84 ± 5%. STAPLE gave the best DC comparison for 79% of all CBCT organ contours. Using one scan takes ∼19 s, fives scans ∼110 s, on a GTX 460 GPU. Combining the plan and all prior days via the STAPLE algorithm to produce the current day contours is superior to the current standard of deforming only the plan contours to the daily CBCT, while maintaining a feasible treatment timescale. Using just the previous day treatment contours provides some improvement. STAPLE also improves the precision, with a substantial decrease in SD, a key for adaptive therapy. There is an immediate gain in accuracy by combining the initial plan with the first day's CBCT (from a second simulation), increasing with more image sets but only marginal improvement beyond the inclusion of 5 CBCTs.