Nicole Schadewaldt

A new method for robust organ positioning in CT images

Robust initialization is essential for any successful segmentation process of medical images. For CT images, initialization is challenging because quality, appearance, content, and field-of-view of the images are highly variable. Furthermore, high execution speed is desirable, whereas the user tolerance to errors is low in clinical applications. We present a new method for efficient and robust positioning of organs in CT images. It is based on a novel probabilistic atlas that, given a tissue type, describes the probability density of the random variable spatial location. Random sampling is then employed to select the most informative points for matching. We present results on pelvic and abdominal images acquired for radiotherapy planning.

SU-E-J-141: Comparison of Dose Calculation On Automatically Generated MRBased ED Maps and Corresponding Patient CT for Clinical Prostate EBRT Plans

PURPOSE To analyze the effect of computing radiation dose on automatically generated MR-based simulated CT images compared to true patient CTs. METHODS Six prostate cancer patients received a regular planning CT for RT planning as well as a conventional 3D fast-field dual-echo scan on a Philips 3.0T Achieva, adding approximately 2 min of scan time to the clinical protocol. Simulated CTs (simCT) where synthesized by assigning known average CT values to the tissue classes air, water, fat, cortical and cancellous bone. For this, Dixon reconstruction of the nearly out-of-phase (echo 1) and in-phase images (echo 2) allowed for water and fat classification. Model based bone segmentation was performed on a combination of the DIXON images. A subsequent automatic threshold divides into cortical and cancellous bone. For validation, the simCT was registered to the true CT and clinical treatment plans were re-computed on the simCT in pinnacle3 . To differentiate effects related to the 5 tissue classes and changes in the patient anatomy not compensated by rigid registration, we also calculate the dose on a stratified CT, where HU values are sorted in to the same 5 tissue classes as the simCT. RESULTS Dose and volume parameters on PTV and risk organs as used for the clinical approval were compared. All deviations are below 1.1%, except the anal sphincter mean dose, which is at most 2.2%, but well below clinical acceptance threshold. Average deviations are below 0.4% for PTV and risk organs and 1.3% for the anal sphincter. The deviations of the stratifiedCT are in the same range as for the simCT. All plans would have passed clinical acceptance thresholds on the simulated CT images. CONCLUSION This study demonstrated the clinical usability of MR based dose calculation with the presented Dixon acquisition and subsequent fully automatic image processing. N. Schadewaldt, H. Schulz, M. Helle and S. Renisch are employed by Phlips Technologie Innovative Techonologies, a subsidiary of Royal Philips NV.

Abdomen segmentation in 3D fetal ultrasound using CNN-powered deformable models

In this paper, voxel probability maps generated by a novel fovea fully convolutional network architecture (FovFCN) are used as additional feature images in the context of a segmentation approach based on deformable shape models. The method is applied to fetal 3D ultrasound image data aiming at a segmentation of the abdominal outline of the fetal torso. This is of interest, e.g., for measuring the fetal abdominal circumference, a standard biometric measure in prenatal screening. The method is trained on 126 3D ultrasound images and tested on 30 additional scans. The results show that the approach can successfully combine the advantages of FovFCNs and deformable shape models in the context of challenging image data, such as given by fetal ultrasound. With a mean error of 2.24 mm, the combination of model-based segmentation and neural networks outperforms the separate approaches.

Annotation-free probabilistic atlas learning for robust anatomy detection in CT images

A fully automatic method generating a whole body atlas from CT images is presented. The atlas serves as a reference space for annotations. It is based on a large collection of partially overlapping medical images and a registration scheme. The atlas itself consists of probabilistic tissue type maps and can represent anatomical variations. The registration scheme is based on an entropy-like measure of these maps and is robust with respect to field-of-view variations. In contrast to other atlas generation methods, which typically rely on a sufficiently large set of annotations on training cases, the presented method requires only the images. An iterative refinement strategy is used to automatically stitch the images to build the atlas. Affine registration of unseen CT images to the probabilistic atlas can be used to transfer reference annotations, e.g. organ models for segmentation initialization or reference bounding boxes for field-of-view selection. The robustness and generality of the method is shown using a three-fold cross-validation of the registration on a set of 316 CT images of unknown content and large anatomical variability. As an example, 17 organs are annotated in the atlas reference space and their localization in the test images is evaluated. The method yields a recall (sensitivity), specificity and precision of at least 96% and thus performs excellent in comparison to competitors.

Automated abdominal plane and circumference estimation in 3D US for fetal screening.

Ultrasound is increasingly becoming a 3D modality. Mechanical and matrix array transducers are able to deliver 3D images with good spatial and temporal resolution. The 3D imaging facilitates the application of automated image analysis to enhance workflows, which has the potential to make ultrasound a less operator dependent modality. However, the analysis of the more complex 3D images and definition of all examination standards on 2D images pose barriers to the use of 3D in daily clinical practice. In this paper, we address a part of the canonical fetal screening program, namely the localization of the abdominal cross-sectional plane with the corresponding measurement of the abdominal circumference in this plane. For this purpose, a fully automated pipeline has been designed starting with a random forest based anatomical landmark detection. A feature trained shape model of the fetal torso including inner organs with the abdominal cross-sectional plane encoded into the model is then transformed into the patient space using the landmark localizations. In a free-form deformation step, the model is individualized to the image, using a torso probability map generated by a convolutional neural network as an additional feature image. After adaptation, the abdominal plane and the abdominal torso contour in that plane are directly obtained. This allows the measurement of the abdominal circumference as well as the rendering of the plane for visual assessment. The method has been trained on 126 and evaluated on 42 abdominal 3D US datasets. An average plane offset error of 5.8 mm and an average relative circumference error of 4.9 % in the evaluation set could be achieved.

Precise anatomy localization in CT data by an improved probabilistic tissue type atlas

Automated interpretation of CT scans is an important, clinically relevant area as the number of such scans is increasing rapidly and the interpretation is time consuming. Anatomy localization is an important prerequisite for any such interpretation task. This can be done by image-to-atlas registration, where the atlas serves as a reference space for annotations such as organ probability maps. Tissue type based atlases allow fast and robust processing of arbitrary CT scans. Here we present two methods which significantly improve organ localization based on tissue types. A first problem is the definition of tissue types, which until now is done heuristically based on experience. We present a method to determine suitable tissue types from sample images automatically. A second problem is the restriction of the transformation space: all prior approaches use global affine maps. We present a hierarchical strategy to refine this global affine map. For each organ or region of interest a localized tissue type atlas is computed and used for a subsequent local affine registration step. A three-fold cross validation on 311 CT images with different fields-of-view demonstrates a reduction of the organ localization error by 33%.

Bladder segmentation in MR images with watershed segmentation and graph cut algorithm

Prostate and cervix cancer diagnosis and treatment planning that is based on MR images benefit from superior soft tissue contrast compared to CT images. For these images an automatic delineation of the prostate or cervix and the organs at risk such as the bladder is highly desirable. This paper describes a method for bladder segmentation that is based on a watershed transform on high image gradient values and gray value valleys together with the classification of watershed regions into bladder contents and tissue by a graph cut algorithm. The obtained results are superior if compared to a simple region-after-region classification.

TU‐G‐134‐03: MR‐Only‐Based Generation of Electron Density Maps and Digitally Reconstructed Radiographs of the Pelvis

PURPOSE To generate electron density (ED) maps and digitally reconstructed radiographs (DRRs) of the pelvis based only on magnetic resonance imaging (MRI). METHODS A conventional 3D fast-field dual-echo sequence was used to acquire image data of 8 healthy subjects with a Philips 3.0T Ingenia TX system in approximately 1:50 min scan time per volunteer. Dixon reconstruction of the nearly out-of-phase (echo 1) and in-phase images (echo 2) allowed for water and fat classification. A bone-enhanced image was generated by automatically thresholding the noise level of the in-phase image with subsequent background removal. ED maps were then produced by assigning known bulk electron densities to the classified bone and tissue fractions. A bone probability atlas derived from CT data was registered to the ED map in order to filter out misclassified voxels. Finally, DRRs were reconstructed from bone-enhanced images as well as from ED maps. RESULTS The proposed MRI sequence with subsequent Dixon reconstruction and probabilistic filtering makes it possible to classify cortical bone, soft tissue and adipose tissue in the pelvis and yields ED maps and corresponding DRRs. Bowel content was misclassified as cortical bone or air and compromised the segmentation in some slices as well as in the DRRs. Automatic probabilistic atlas filtering can significantly reduce artifacts induced by bowel content without affecting pelvic bone structures markedly. In total, the artifact/bone fraction dropped from 1.7 before filtering to 0.2 after filtering. The average reduction of artifact volume is 87%, and the average bone preservation is 99%. Remaining artifacts are spatially close to the true bone in areas of positive bone probability. CONCLUSION This study demonstrated the feasibility of generating realistic ED maps of the pelvis by using MRI only. The method has the potential to become an essential component of emerging applications such as MR-only-based radiation therapy planning. All authors have the following relevant financial interest or relationship to disclose with regard to the subject matter of this presentation: Company name: Philips Research; Type of relationship: Employee.

A completely automated processing pipeline for lung and lung lobe segmentation and its application to the LIDC-IDRI data base

Automated segmentation of lung lobes in thoracic CT images has relevance for various diagnostic purposes like localization of tumors within the lung or quantification of emphysema. Since emphysema is a known risk factor for lung cancer, both purposes are even related to each other. The main steps of the segmentation pipeline described in this paper are the lung detector and the lung segmentation based on a watershed algorithm, and the lung lobe segmentation based on mesh model adaptation. The segmentation procedure was applied to data sets of the data base of the Image Database Resource Initiative (IDRI) that currently contains over 500 thoracic CT scans with delineated lung nodule annotations. We visually assessed the reliability of the single segmentation steps, with a success rate of 98% for the lung detection and 90% for lung delineation. For about 20% of the cases we found the lobe segmentation not to be anatomically plausible. A modeling confidence measure is introduced that gives a quantitative indication of the segmentation quality. For a demonstration of the segmentation method we studied the correlation between emphysema score and malignancy on a per-lobe basis.