Benjamin Haas

Automatic segmentation of thoracic and pelvic CT images for radiotherapy planning using implicit anatomic knowledge and organ-specific segmentation strategies

Automatic segmentation of anatomical structures in medical images is a valuable tool for efficient computer-aided radiotherapy and surgery planning and an enabling technology for dynamic adaptive radiotherapy. This paper presents the design, algorithms and validation of new software for the automatic segmentation of CT images used for radiotherapy treatment planning. A coarse to fine approach is followed that consists of presegmentation, anatomic orientation and structure segmentation. No user input or a priori information about the image content is required. In presegmentation, the body outline, the bones and lung equivalent tissue are detected. Anatomic orientation recognizes the patients position, orientation and gender and creates an elastic mapping of the slice positions to a reference scale. Structure segmentation is divided into localization, outlining and refinement, performed by procedures with implicit anatomic knowledge using standard image processing operations. The presented version of algorithms automatically segments the body outline and bones in any gender and patient position, the prostate, bladder and femoral heads for male pelvis in supine position, and the spinal canal, lungs, heart and trachea in supine position. The software was developed and tested on a collection of over 600 clinical radiotherapy planning CT stacks. In a qualitative validation on this test collection, anatomic orientation correctly detected gender, patient position and body region in 98% of the cases, a correct mapping was produced for 89% of thorax and 94% of pelvis cases. The average processing time for the entire segmentation of a CT stack was less than 1 min on a standard personal computer. Two independent retrospective studies were carried out for clinical validation. Study I was performed on 66 cases (30 pelvis, 36 thorax) with dosimetrists, study II on 52 cases (39 pelvis, 13 thorax) with radio-oncologists as experts. The experts rated the automatically produced structures on the scale 1-excellent (no corrections necessary, maximum time saving), 2-good (corrections necessary for up to 1/3 of slices), 3-acceptable (major corrections necessary, but still time saving), 4-not acceptable (manual redrawing more efficient, no time saving). A rating<or=3 indicates a time saving in the treatment planning process and was given for pelvis segmentation in 70% (I) and 68% (II) of the cases, with average ratings 2.9 (I) and 2.6 (II). For the thorax, a rating<or=3 was given in 94% and 91% of the cases, with average ratings 2.1 and 1.9, respectively. For quantitative validation, automatically generated structures were compared geometrically in 2D and 3D to manually drawn structures created by experts on seven randomly selected cases. The quantitative validation was limited to pelvis structures. The results indicate that the accuracy of the algorithms is within the bandwidth of manual segmentation by experts, except for specific erroneous situations. Even though manual review and corrections of automatically segmented structures are still mandatory, it can be expected that due to the speed of the presented software and the quality of its results, its introduction in the radiotherapy treatment planning process will lead to a considerable amount of time being saved.

Deformable image registration for adaptive radiotherapy with guaranteed local rigidity constraints

BackgroundDeformable image registration (DIR) is a key component in many radiotherapy applications. However, often resulting deformations are not satisfying, since varying deformation properties of different anatomical regions are not considered. To improve the plausibility of DIR in adaptive radiotherapy in the male pelvic area, this work integrates a local rigidity deformation model into a DIR algorithm.MethodsA DIR framework is extended by constraints, enforcing locally rigid deformation behavior for arbitrary delineated structures. The approach restricts those structures to rigid deformations, while surrounding tissue is still allowed to deform elastically. The algorithm is tested on ten CT/CBCT male pelvis datasets with active rigidity constraints on bones and prostate and compared to the Varian SmartAdapt deformable registration (VSA) on delineations of bladder, prostate and bones.ResultsThe approach with no rigid structures (REG0) obtains an average dice similarity coefficient (DSC) of 0.87 ± 0.06 and a Hausdorff-Distance (HD) of 8.74 ± 5.95 mm. The new approach with rigid bones (REG1) yields a DSC of 0.87 ± 0.07, HD 8.91 ± 5.89 mm. Rigid deformation of bones and prostate (REG2) obtains 0.87 ± 0.06, HD 8.73 ± 6.01 mm, while VSA yields a DSC of 0.86 ± 0.07, HD 10.22 ± 6.62 mm. No deformation grid foldings are observed for REG0 and REG1 in 7 of 10 cases; for REG2 in 8 of 10 cases, with no grid foldings in prostate, an average of 0.08 % in bladder (REG2: no foldings) and 0.01 % inside the body contour. VSA exhibits grid foldings in each case, with an average percentage of 1.81 % for prostate, 1.74 % for bladder and 0.12 % for the body contour. While REG1 and REG2 keep bones rigid, elastic bone deformations are observed with REG0 and VSA. An average runtime of 26.2 s was achieved with REG1; 31.1 s with REG2, compared to 10.5 s with REG0 and 10.7 s with VMS.ConclusionsWith accuracy in the range of VSA, the new approach with constraints delivers physically more plausible deformations in the pelvic area with guaranteed rigidity of arbitrary structures. Although the algorithm uses an advanced deformation model, clinically feasible runtimes are achieved.

Comparison of different deep learning approaches for parotid gland segmentation from CT images

The segmentation of target structures and organs at risk is a crucial and very time-consuming step in radiotherapy planning. Good automatic methods can significantly reduce the time clinicians have to spend on this task. Due to its variability in shape and often low contrast to surrounding structures, segmentation of the parotid gland is especially challenging. Motivated by the recent success of deep learning, we study different deep learning approaches for parotid gland segmentation. Particularly, we compare 2D, 2D ensemble and 3D U-Net approaches and find that the 2D U-Net ensemble yields the best results with a mean Dice score of 0.817 on our test data. The ensemble approach reduces false positives without the need for an automatic region of interest detection. We also apply our trained 2D U-Net ensemble to segment the test data of the 2015 MICCAI head and neck auto-segmentation challenge. With a mean Dice score of 0.861, our classifier exceeds the highest mean score in the challenge. This shows that the method generalizes well onto data from independent sites. Since appropriate reference annotations are essential for training but often difficult and expensive to obtain, it is important to know how many samples are needed to properly train a neural network. We evaluate the classifier performance after training with differently sized training sets (50–450) and find that 250 cases (without using extensive data augmentation) are sufficient to obtain good results with the 2D ensemble. Adding more samples does not significantly improve the Dice score of the segmentations.

Atlas ranking and selection for automatic segmentation of the esophagus from CT scans

In radiation treatment planning, the esophagus is an important organ-at-risk that should be spared in patients with head and neck cancer or thoracic cancer who undergo intensity-modulated radiation therapy. However, automatic segmentation of the esophagus from CT scans is extremely challenging because of the structures inconsistent intensity, low contrast against the surrounding tissues, complex and variable shape and location, and random air bubbles. The goal of this study is to develop an online atlas selection approach to choose a subset of optimal atlases for multi-atlas segmentation to the delineate esophagus automatically. We performed atlas selection in two phases. In the first phase, we used the correlation coefficient of the image content in a cubic region between each atlas and the new image to evaluate their similarity and to rank the atlases in an atlas pool. A subset of atlases based on this ranking was selected, and deformable image registration was performed to generate deformed contours and deformed images in the new image space. In the second phase of atlas selection, we used Kullback-Leibler divergence to measure the similarity of local-intensity histograms between the new image and each of the deformed images, and the measurements were used to rank the previously selected atlases. Deformed contours were overlapped sequentially, from the most to the least similar, and the overlap ratio was examined. We further identified a subset of optimal atlases by analyzing the variation of the overlap ratio versus the number of atlases. The deformed contours from these optimal atlases were fused together using a modified simultaneous truth and performance level estimation algorithm to produce the final segmentation. The approach was validated with promising results using both internal data sets (21 head and neck cancer patients and 15 thoracic cancer patients) and external data sets (30 thoracic patients).

Accuracy of patient specific organ-dose estimates obtained using an automated image segmentation algorithm

The overall goal of this work is to develop a rapid, accurate and fully automated software tool to estimate patient-specific organ doses from computed tomography (CT) scans using a deterministic Boltzmann Transport Equation solver and automated CT segmentation algorithms. This work quantified the accuracy of organ dose estimates obtained by an automated segmentation algorithm. The investigated algorithm uses a combination of feature-based and atlas-based methods. A multiatlas approach was also investigated. We hypothesize that the auto-segmentation algorithm is sufficiently accurate to provide organ dose estimates since random errors at the organ boundaries will average out when computing the total organ dose. To test this hypothesis, twenty head-neck CT scans were expertly segmented into nine regions. A leave-one-out validation study was performed, where every case was automatically segmented with each of the remaining cases used as the expert atlas, resulting in nineteen automated segmentations for each of the twenty datasets. The segmented regions were applied to gold-standard Monte Carlo dose maps to estimate mean and peak organ doses. The results demonstrated that the fully automated segmentation algorithm estimated the mean organ dose to within 10% of the expert segmentation for regions other than the spinal canal, with median error for each organ region below 2%. In the spinal canal region, the median error was 7% across all data sets and atlases, with a maximum error of 20%. The error in peak organ dose was below 10% for all regions, with a median error below 4% for all organ regions. The multiple-case atlas reduced the variation in the dose estimates and additional improvements may be possible with more robust multi-atlas approaches. Overall, the results support potential feasibility of an automated segmentation algorithm to provide accurate organ dose estimates.

Accuracy of Patient-Specific Organ Dose Estimates Obtained Using an Automated Image Segmentation Algorithm

Abstract. The overall goal of this work is to develop a rapid, accurate, and automated software tool to estimate patient-specific organ doses from computed tomography (CT) scans using simulations to generate dose maps combined with automated segmentation algorithms. This work quantified the accuracy of organ dose estimates obtained by an automated segmentation algorithm. We hypothesized that the autosegmentation algorithm is sufficiently accurate to provide organ dose estimates, since small errors delineating organ boundaries will have minimal effect when computing mean organ dose. A leave-one-out validation study of the automated algorithm was performed with 20 head-neck CT scans expertly segmented into nine regions. Mean organ doses of the automatically and expertly segmented regions were computed from Monte Carlo-generated dose maps and compared. The automated segmentation algorithm estimated the mean organ dose to be within 10% of the expert segmentation for regions other than the spinal canal, with the median error for each organ region below 2%. In the spinal canal region, the median error was −7%, with a maximum absolute error of 28% for the single-atlas approach and 11% for the multiatlas approach. The results demonstrate that the automated segmentation algorithm can provide accurate organ dose estimates despite some segmentation errors.