T Zhao

Two‐stage atlas subset selection in multi‐atlas based image segmentation

PURPOSE Fast growing access to large databases and cloud stored data presents a unique opportunity for multi-atlas based image segmentation and also presents challenges in heterogeneous atlas quality and computation burden. This work aims to develop a novel two-stage method tailored to the special needs in the face of large atlas collection with varied quality, so that high-accuracy segmentation can be achieved with low computational cost. METHODS An atlas subset selection scheme is proposed to substitute a significant portion of the computationally expensive full-fledged registration in the conventional scheme with a low-cost alternative. More specifically, the authors introduce a two-stage atlas subset selection method. In the first stage, an augmented subset is obtained based on a low-cost registration configuration and a preliminary relevance metric; in the second stage, the subset is further narrowed down to a fusion set of desired size, based on full-fledged registration and a refined relevance metric. An inference model is developed to characterize the relationship between the preliminary and refined relevance metrics, and a proper augmented subset size is derived to ensure that the desired atlases survive the preliminary selection with high probability. RESULTS The performance of the proposed scheme has been assessed with cross validation based on two clinical datasets consisting of manually segmented prostate and brain magnetic resonance images, respectively. The proposed scheme demonstrates comparable end-to-end segmentation performance as the conventional single-stage selection method, but with significant computation reduction. Compared with the alternative computation reduction method, their scheme improves the mean and medium Dice similarity coefficient value from (0.74, 0.78) to (0.83, 0.85) and from (0.82, 0.84) to (0.95, 0.95) for prostate and corpus callosum segmentation, respectively, with statistical significance. CONCLUSIONS The authors have developed a novel two-stage atlas subset selection scheme for multi-atlas based segmentation. It achieves good segmentation accuracy with significantly reduced computation cost, making it a suitable configuration in the presence of extensive heterogeneous atlases.

TH-CD-206-04: Learning Relevance Criterion for Multi-Atlas Based Image Segmentation

PURPOSE It is important to effectively identify the subset of relevant training atlases in learning based segmentation. Since segmentation based geometry relevance between atlases and the target is inaccessible due to the unknown target segmentation, image feature based relevance criteria have to be used instead. This study aims to learn an image based criterion that best reflects the underlying geometry agreement. METHODS We learn the image-based relevance criterion in the form of a Mahalanobis distance on linearly transformed image feature, which is specialized to image intensity in this work, and optimize such linear transformation. Given a set of atlases, i.e., image/segmentation pairs, a linear transformation is optimized to make the Mahalanobis distance small between the geometrically relevant atlases and large between the irrelevant pairs. Performance assessment and comparison with the commonly-used mean square distance (MSD), a special case of the Mahalanobis distance with a trivial linear transform, are performed based on clinical brain MR images. RESULTS The proposed surrogate learning approach was validated with multi-atlas based corpus callosum segmentation. Compared to MSD, our learned surrogate demonstrated superiority in selecting the geometrically relevant atlases. The learned surrogate yielded an improvement on the average pairwise Dice similarity coefficient (DSC) of the selected atlases from .004 to .007, and a corresponding improvement on the ultimate segmentation accuracy from .003 to .005, over an atlas subset size from 1 to 20. CONCLUSION This work provides a systematic methodology to learn task-specific criterion to select atlases for image segmentation, and demonstrates effectiveness in identifying the most relevant atlases. We are working on investigating various image features and extending to nonlinear transformations for further improvement.

Image segmentation with a novel regularized composite shape prior based on surrogate study

PURPOSE Incorporating training into image segmentation is a good approach to achieve additional robustness. This work aims to develop an effective strategy to utilize shape prior knowledge, so that the segmentation label evolution can be driven toward the desired global optimum. METHODS In the variational image segmentation framework, a regularization for the composite shape prior is designed to incorporate the geometric relevance of individual training data to the target, which is inferred by an image-based surrogate relevance metric. Specifically, this regularization is imposed on the linear weights of composite shapes and serves as a hyperprior. The overall problem is formulated in a unified optimization setting and a variational block-descent algorithm is derived. RESULTS The performance of the proposed scheme is assessed in both corpus callosum segmentation from an MR image set and clavicle segmentation based on CT images. The resulted shape composition provides a proper preference for the geometrically relevant training data. A paired Wilcoxon signed rank test demonstrates statistically significant improvement of image segmentation accuracy, when compared to multiatlas label fusion method and three other benchmark active contour schemes. CONCLUSIONS This work has developed a novel composite shape prior regularization, which achieves superior segmentation performance than typical benchmark schemes.

TU-CD-BRA-05: Atlas Selection for Multi-Atlas-Based Image Segmentation Using Surrogate Modeling

Purpose: The growing size and heterogeneity in training atlas necessitates sophisticated schemes to identify only the most relevant atlases for the specific multi-atlas-based image segmentation problem. This study aims to develop a model to infer the inaccessible oracle geometric relevance metric from surrogate image similarity metrics, and based on such model, provide guidance to atlas selection in multi-atlas-based image segmentation. Methods: We relate the oracle geometric relevance metric in label space to the surrogate metric in image space, by a monotonically non-decreasing function with additive random perturbations. Subsequently, a surrogate’s ability to prognosticate the oracle order for atlas subset selection is quantified probabilistically. Finally, important insights and guidance are provided for the design of fusion set size, balancing the competing demands to include the most relevant atlases and to exclude the most irrelevant ones. A systematic solution is derived based on an optimization framework. Model verification and performance assessment is performed based on clinical prostate MR images. Results: The proposed surrogate model was exemplified by a linear map with normally distributed perturbation, and verified with several commonly-used surrogates, including MSD, NCC and (N)MI. The derived behaviors of different surrogates in atlas selection and their corresponding performance in ultimate label estimate were validated. The performance of NCC and (N)MI was similarly superior to MSD, with a 10% higher atlas selection probability and a segmentation performance increase in DSC by 0.10 with the first and third quartiles of (0.83, 0.89), compared to (0.81, 0.89). The derived optimal fusion set size, valued at 7/8/8/7 for MSD/NCC/MI/NMI, agreed well with the appropriate range [4, 9] from empirical observation. Conclusion: This work has developed an efficacious probabilistic model to characterize the image-based surrogate metric on atlas selection. Analytical insights lead to valid guiding principles on fusion set size design.

SU‐E‐J‐136: Evaluation of a Non‐Invasive Method on Lung Tumor Tracking

PURPOSE to develop a non-invasive method to track lung motion in free-breathing patients. METHODS A free-breathing breathing model has been developed to use tidal volume and air flow rate as surrogates for lung trajectories. In this study, 4D CT data sets were acquired during simulation and were reconstructed into 10 phases. Total lung capacities were calculated from the reconstructed images. Continuous signals from the abdominal pneumatic belt were correlated to the volumes and were therefore converted into a curve of tidal volumes. Air flow rate were calculated as the first order derivative of the tidal volume curve. Lung trajectories in the 10 reconstructed images were obtained using B-Spline registration. Parameters of the free-breathing lung motion model were fit from the tidal volumes, airflow rates and lung trajectories using the simulation data. Patients were rescanned every week during the treatment. Prediction of lung trajectories from the model were given and compared to the actual positions in BEV. RESULTS Trajectories of lung were predicted with residual error of 1.49mm at 95th percentile of all tracked points. Tracking was stable and reproducible over two weeks. CONCLUSION Non-invasive tumor tracking based on a free-breathing lung motion model is feasible and stable over weeks.

SU‐E‐J‐107: Utilization of Hysteresis Motion as a Marker for Tumor Margin in a Free‐Breathing Lung

Purpose: To study the feasibility of using hysteresis motion characterized by a free‐breathing lung motion model as a marker for tumor margin in lung patients. Methods: Anatomical and physiological evidences have shown that solid tumors typically have more collagen in tumor stroma and consequently are firmer than surrounding normal tissues. For patients with lungcancer, the increased firmness in tumor stroma would alter pattern of breathing motion, mostly prominent in the hysteresis component that is defined as the variation between motion trajectories during inhalation and exhalation. A free‐breathing lung motion model, x= x0+alpha*v+beta*f, where v and f denote tidal volume and air flow respectively, decomposes breathing motion into a non‐hysteresis component alpha*v, which is purely due to air filling, and a hysteresis component beta*f. Four patients acquired in Cine mode were analyzed and beta was obtained by linear least‐square fitting registered motions to air flows measured by spirometry. Results: Hysteresis motions were found to gently change in magnitude and direction throughout whole lungs except around cancerous regions where the hysteresis motions turn sharply following the curvature of the tumors, indicating shear motions along the interface between tumors and surrounding normal tissues. Conclusions: The vector map of hysteresis motion demonstrates affinity to tumor curvature in the vicinity of the tumor, reflecting intrinsic physiological and mechanical changes at the transition from tumor stroma to surrounding normal tissue. It has the potential to be used as a marker for tumor margin.

SU‐E‐T‐511: Distribution of Hysteresis Magnitude during Free Breathing

Purpose: To characterize and quantify the hysteresis motion component distribution during free breathing. Methods: We have hypothesized that lung tissue motion can be represented as a linear combination of two components: due to volume filling and due to hysteresis. Lung tissue trajectories were calculated using an existing 5D lung tissue trajectory model that uses two vector fields, alpha and beta, which were previously determined. Physically, the alpha motion vector was the tissue‐specific volume filling component and beta was the tissue‐specific hysteresis component. Due to natural variations in breathing patterns, we developed the concept of a characteristic breath. Each breath was corrected with a linear drift model from exhalation to exhalation. In order to characterize the hysteresis motion, the ratio of the volume filling to hysteresis components was examined throughout the subjects lungs by generating a bounding box with one side parallel to the alpha vector and the box lying in the plane defined by the alpha and beta vectors. The lungs were subdivided into geographic quadrants to observe the variation of the bounding box elongation. Because the displacement varied, the 15th and 85th percentile tidal volume displacements were selected to define the range of analyzed displacements. This study utilized 50 subject data sets from a 16 slice CT scanner. Results: The 15th and 85th percentile bounding box elongations were 0.090±0.005 and 0.187±0.037, and 0.083±0.013 and 0.203±0.053, in the upper and lower left lung and 0.092±0.006 and 0.184±0.038, and 0.085±0.013 and 0.196±0.043, in the upper and lower right lung, respectively, for volume displacements between 5–15mm. Conclusion: Hysteresis motion was relatively small compared to the volume‐filling motion, contributing between 8% and 20% of the overall motion. Little difference in this range was observed for upper and lower lung regions. This work supported in part by NIHR01CA116712 and NIHR01CA96679.

SU‐GG‐J‐90: Quantification of the Thorax‐To‐Abdomen Breathing Ratio for Breathing Motion Modeling

Purpose: To develop a methodology to quantitatively measure the thorax‐to‐abdomen breathing ratio for breathing motion modeling and breathing motion studies. Method and Materials: The breathing ratio was quantified by measuring the rate of body volume increase throughout the thorax and abdomen as a function of tidal volume. 15 16‐slice 4DCT patient imagedata sets from the neck through the pelvis were acquired during quiet respiration using a protocol that acquired 25 cine scans at each couch position. Tidal volume was used as the breathing‐cycle surrogate, measured using a spirometer and abdominal pneumatic bellows. The volume within the skin contour at each CT slice was compared against the tidal volume, exhibiting a nearly linear relationship. A linear regression analysis was used to determine η(i), defined as the amount of expansion at each slice i per unit tidal volume. The sum Ση(i) throughout all slices was predicted to be the ratio of room air density to internal air density; 1.11. The boundary between the thorax and abdomen was determined by examining the patient anatomy and setting the boundary at the Xiphoid process. The thorax and abdomen regions were individually analyzed to determine the thorax‐to‐abdomen breathing ratios. Results: The average Ση(i) for all data sets was found to be 1.23±0.20, close to the expected value of 1.11. The thorax‐to‐abdomen breathing ratios were 0.22±0.19. The average Ση(i) was 0.20±0.13 in the thorax and 1.04±0.26 in the abdomen. The boundary between the thorax and abdomen was localized using a 50th tidal volume percentile reconstructedCTimage.Conclusion: A method to quantify the relationship between abdomen and thoracic breathing was developed and validated. This work supported in part by NIHR01CA116712 and NIHR01CA96679.