Segmentation for Radiotherapy Planning via M-reps MIDAG - Medical Image Display and Analysis Group

Overview of Evaluation Study

Segmentation of volume images, e.g., CT and MR images, is a common clinical procedure required for defining normal anatomical structures and tumor volumes for 3D radiation therapy treatment planning (3D RTP). Current user-guided methods are inefficient and exhibit significant inter-and intra-user variability. Improvements in efficiency can reduce costs and possibly facilitate new technologies for target localization and patient positioning immediately prior to each treatment fraction and for online verification of treatment geometry using electronic portal images. Improved accuracy and precision is clinically important in many cases, particularly when the risk of morbidity is a significant factor in treatment planning decisions, and for clinical trials that collect dose-volume data that will be factored into determination of probabilities for tumor control and normal tissue complications. We are in the early stages of an evaluation of the accuracy, reproducibility, robustness and sensitivity of M-rep segmentation for selected normal structures important in 3D RTP. Our general procedure is to build and train M-rep models of anatomical objects, apply the models to target images for which truth is defined, and measure the goodness of match between the deformed models and truth. Target images are samples from a disribution derived from a labeled reference image by applying warp fields that map original intensity values to new locations. The warp fields change the shapes of objects in the reference image in a way that mimics normal shape variations. Noise and other image disturbances will be added to target images in a graded manner to test robustness and sensitivity. Image parameters to be investigated include modality, slice thickness, contrast vs non contrast, and MR pulse sequence. M-rep characteristics to be investigated include model builder preferences (e.g., different model building decisions that result in structurally distinct models); medial atom density and spacing; atom-level choices for image sampling filters; hybrid models (e.g., models including landmarks or parts of multiple objects); and linked M-reps for multiple objects (local atlases).

Geometric Evaluation

Geometric scoring will include comparison of reference and segmented volumes; quantification of union voxels, excluded true voxels, included false voxels, and mean and maximum displacement of surfaces; and analysis for systematic and random shape differences.

Clinical Evaluation

We plan to compare M-rep and user-guided segmentation in the clinical context of forward and inverse treatment planning. Our approach for forward planning will involve resegmenting archived planning images, and recomputing dose-volume histograms using the new M-rep segmentations. New and historic dose-volume histograms will be analyzed for clinically significant differences that might have affected treatment planning decisions. For inverse planning we will investigate algorithm-driven dose optimization using planning images that have been segmented with conventional and M-rep approaches. Inverse planning algorithms yield modulated treatment beams with strong intensity gradients near tumor and normal tissue boundaries, and thus these algorithms are exquisitely sensitive to segmentation results.

Grant Support

NCI P01 CA47982


Mills PH, Fuchs H, Pizer SM, Rosenman JG, IMEX: A tool for image display and contour management in a windowing environment. Medical Imaging III: Image Capture and Display, RH Schneider, S Dwyer II, and RG Jost, eds, Proceedings of the SPIE 1091:132-142, 1989

Chaney E and Pizer S, Defining Anatomical Structures from Medical Images, Seminars in Radiation Oncology 2(4):215-225 (1992)

Tracton G, Chaney EL, Rosenman JG, Pizer SM, MASK: Combining 2D and 3D Segmentation Methods to Enhance Functionality, Math. Methods Med. Imaging III, F.L. Bookstein et al eds, SPIE Proc. Volume 2299: 98-109 (1994)

McAuliffe MJ, Eberly D, Fritsch DS, Chaney EL, and Pizer SM: Scale-Space Boundary Evolution Initialized by Cores. Proceedings of the Fourth International Conference on Visualization in Biomedical Computing, Lecture Notes in Comp Sci Vol 1131, KH Hohne and R Kikinis (eds), Springer: 173-182 (1996)

Fritsch DS, Pizer SM, Yu L, Johnson V, and Chaney EL: Segmentation of Medical Image Objects using Deformable Shape Loci, Proc. Info. Processing in Med. Imaging (IPMI), Lecture Notes in Computer Science, J Duncan and G Gindi, eds, Springer-Verlag, Berlin: 127-140 (1997)

Chaney EL, Fritsch DS, Pizer S, Objective GTV and Normal Structure Segmentation, Imaging in Radiotherapy, A. Boyer and J. Hazle, eds., Advanced Medical Publishing, Madison, WI: 271-290 (1998)

Pizer SM, Fritsch DS, Yushkevich PA, Johnson VE, and Chaney EL: Segmentation, Registration, and Measurement of Shape Variation via Image Object Shape, Transactions in Medical Imaging 18(10): 851-865 (1999)