Segmenting CT prostate images using population and patient-specific statistics for radiotherapy
Comparison between intra- and interpatient variations. (a) Prostate shapes of three patients at four time points. Shapes in each row are from the same patient; shapes in different columns are from different time points. (b) The ASDs among 24 patients. For one patient, the intrapatient ASD is the mean of distances between each shape and the mean shape of that patient, and the interpatient ASD is the distance between the mean shape of that patient and the total mean shape of all patients.
Flow diagram of template-based surface construction framework. The arrows with light color (blue in online version) represent the surface deformation procedures using the method proposed in Sec. ???.
Demonstration of the map of the prostate. In this figure, a prostate surface was color coded according to the value at each vertex.
Demonstration of profile truncation. The profiles which lie across the regions associated with gas or bone are truncated. In this figure, there are two profiles truncated to an appropriate length.
Typical results illustrate the effectiveness of the optimal length profile. The dark (red in online version) contours show the hand-drawn ground-truth contours, and the light (green in online version) and gray (blue in online version) contours represent the results using the fixed length profiles and the proposed optimal length profiles, respectively. It is clear that the segmentation accuracy is significantly improved for the boundary segment adjacent to the rectum in (a) and (b), as well as for the boundary segment adjacent to the bone in (b) by using the proposed optimal length profiles.
Flow diagram of shape model online training. In this figure, the blocks with dark color (green in online version) represent the population information, and the light (yellow in online version) one represents the patient-specific information.
Illustration of the values of parameter at different time points. In this figure, after the number of captured images of the current patient reaches 9, it approximately equals 1. That implies that patient-specific information totally controls the shape model after that time.
DSCs (a) and ASDs (b) of all 306 images by using gradient and gradient-PDF combination features, respectively. The image order is ranked according to the DSC values of segmentation results using gradient-PDF combination features for better illustration.
Segmentation results for evenly spaced slices of image 8 of patient 24 (DSC of 92.2%). The light (yellow in online version) contours show the result of the proposed method. The dark (red in online version) contours show the hand-drawn (ground-truth) contours supplied by a radiation oncologist.
Average DSCs of 24 patients at each treatment day (1–9). Dotted lines show the trends of segmentation performance.
Average DSC and average ASD between the manual and the automated segmentations of all 306 images using gradient features and gradient-PDF combination features, respectively.
Average DSC and average ASD between the manual and the automated segmentations of all 306 images under three different update strategies, respectively.
Article metrics loading...
Full text loading...