Automation of ultrasonographic optic nerve sheath diameter measurement using convolutional neural networks.

BACKGROUND AND PURPOSE: Ultrasonographic optic nerve sheath (ONS) diameter is a noninvasive intracranial pressure (ICP) surrogate. ICP is monitored invasively in specialized intensive care units. Noninvasive ICP monitoring is important in less specialized settings. However, noninvasive ICP monitoring using ONS diameter (ONSD) is limited by the need for experts to obtain and perform measurements. We aim to automate ONSD measurements using a deep convolutional neural network (CNN) with a novel masking technique. METHODS: We trained a CNN to reproduce masks that mark the ONS. The edges of the mask are defined by an expert. Eight models were trained with 1000 epochs per model. The Dice-similarity-coefficient-weighted averaged outputs of the eight models yielded the final predicted mask. Eight hundred and seventy-three images were obtained from 52 transorbital cine-ultrasonography sessions, performed on 46 patients with brain injuries. Eight hundred and fourteen images from 48 scanning sessions were used for training and validation and 59 images from four sessions for testing. Bland-Altman and Pearson linear correlation analyses were used to evaluate the agreement between CNN and expert measurements. RESULTS: Expert ONSD measurements and CNN-derived ONSD estimates had strong agreement (r = 0.7, p < .0001). The expert mean ONSD (standard deviation) is 5.27 mm (0.43) compared to CNN mean estimate of 5.46 mm (0.37). Mean difference (95% confidence interval, p value) is 0.19 mm (0.10-0.27 mm, p = .0011), and root mean square error is 0.27 mm. CONCLUSION: A CNN can learn ONSD measurement using masking without image segmentation or landmark detection.

Automated Tumor Segmentation in Radiotherapy.

Autosegmentation of gross tumor volumes holds promise to decrease clinical demand and to provide consistency across clinicians and institutions for radiation treatment planning. Additionally, autosegmentation can enable imaging analyses such as radiomics to construct and deploy large studies with thousands of patients. Here, we review modern results that utilize deep learning approaches to segment tumors in 5 major clinical sites: brain, head and neck, thorax, abdomen, and pelvis. We focus on approaches that inch closer to clinical adoption, highlighting winning entries in international competitions, unique network architectures, and novel ways of overcoming specific challenges. We also broadly discuss the future of gross tumor volumes autosegmentation and the remaining barriers that must be overcome before widespread replacement or augmentation of manual contouring.

Pediatric chest-abdomen-pelvis and abdomen-pelvis CT images with expert organ contours.

PURPOSE: Organ autosegmentation efforts to date have largely been focused on adult populations, due to limited availability of pediatric training data. Pediatric patients may present additional challenges for organ segmentation. This paper describes a dataset of 359 pediatric chest-abdomen-pelvis and abdomen-pelvis Computed Tomography (CT) images with expert contours of up to 29 anatomical organ structures to aid in the evaluation and development of autosegmentation algorithms for pediatric CT imaging. ACQUISITION AND VALIDATION METHODS: The dataset collection consists of axial CT images in Digital Imaging and Communications in Medicine (DICOM) format of 180 male and 179 female pediatric chest-abdomen-pelvis or abdomen-pelvis exams acquired from one of three CT scanners at Children's Wisconsin. The datasets represent random pediatric cases based upon routine clinical indications. Subjects ranged in age from 5 days to 16 years, with a mean age of 7 years. The CT acquisition, contrast, and reconstruction protocols varied across the scanner models and patients, with specifications available in the DICOM headers. Expert contours were manually labeled for up to 29 organ structures per subject. Not all contours are available for all subjects, due to limited field of view or unreliable contouring due to high noise. DATA FORMAT AND USAGE NOTES: The data are available on The Cancer Imaging Archive (TCIA_ (https://www.cancerimagingarchive.net/) under the collection Pediatric-CT-SEG. The axial CT image slices for each subject are available in DICOM format. The expert contours are stored in a single DICOM RTSTRUCT file for each subject. The contour names are listed in Table 2. POTENTIAL APPLICATIONS: This dataset will enable the evaluation and development of organ autosegmentation algorithms for pediatric populations, which exhibit variations in organ shape and size across age. Automated organ segmentation from CT images has numerous applications including radiation therapy, diagnostic tasks, surgical planning, and patient-specific organ dose estimation.

Beam-centric algorithm for pretreatment patient position correction in external beam radiation therapy.

PURPOSE: In current image guided pretreatment patient position adjustment methods, image registration is used to determine alignment parameters. Since most positioning hardware lacks the full six degrees of freedom (DOF), accuracy is compromised. The authors show that such compromises are often unnecessary when one models the planned treatment beams as part of the adjustment calculation process. The authors present a flexible algorithm for determining optimal realizable adjustments for both step-and-shoot and arc delivery methods. METHODS: The beam shape model is based on the polygonal intersection of each beam segment with the plane in pretreatment image volume that passes through machine isocenter perpendicular to the central axis of the beam. Under a virtual six-DOF correction, ideal positions of these polygon vertices are computed. The proposed method determines the couch, gantry, and collimator adjustments that minimize the total mismatch of all vertices over all segments with respect to their ideal positions. Using this geometric error metric as a function of the number of available DOF, the user may select the most desirable correction regime. RESULTS: For a simulated treatment plan consisting of three equally weighted coplanar fixed beams, the authors achieve a 7% residual geometric error (with respect to the ideal correction, considered 0% error) by applying gantry rotation as well as translation and isocentric rotation of the couch. For a clinical head-and-neck intensity modulated radiotherapy plan with seven beams and five segments per beam, the corresponding error is 6%. Correction involving only couch translation (typical clinical practice) leads to a much larger 18% mismatch. Clinically significant consequences of more accurate adjustment are apparent in the dose volume histograms of target and critical structures. CONCLUSIONS: The algorithm achieves improvements in delivery accuracy using standard delivery hardware without significantly increasing total treatment session duration. It encourages parsimonious utilization of all available DOF. Finally, in certain cases, it obviates the need of a robotic couch having six DOF for the correction of patient displacement and rotations.

Algorithm for X-ray scatter, beam-hardening, and beam profile correction in diagnostic (kilovoltage) and treatment (megavoltage) cone beam CT.

Quantitative reconstruction of cone beam X-ray computed tomography (CT) datasets requires accurate modeling of scatter, beam-hardening, beam profile, and detector response. Typically, commercial imaging systems use fast empirical corrections that are designed to reduce visible artifacts due to incomplete modeling of the image formation process. In contrast, Monte Carlo (MC) methods are much more accurate but are relatively slow. Scatter kernel superposition (SKS) methods offer a balance between accuracy and computational practicality. We show how a single SKS algorithm can be employed to correct both kilovoltage (kV) energy (diagnostic) and megavoltage (MV) energy (treatment) X-ray images. Using MC models of kV and MV imaging systems, we map intensities recorded on an amorphous silicon flat panel detector to water-equivalent thicknesses (WETs). Scattergrams are derived from acquired projection images using scatter kernels indexed by the local WET values and are then iteratively refined using a scatter magnitude bounding scheme that allows the algorithm to accommodate the very high scatter-to-primary ratios encountered in kV imaging. The algorithm recovers radiological thicknesses to within 9% of the true value at both kV and megavolt energies. Nonuniformity in CT reconstructions of homogeneous phantoms is reduced by an average of 76% over a wide range of beam energies and phantom geometries.

Automatic coregistration of volumetric images based on implanted fiducial markers.

The accurate delivery of external beam radiation therapy is often facilitated through the implantation of radio-opaque fiducial markers (gold seeds). Before the delivery of each treatment fraction, seed positions can be determined via low dose volumetric imaging. By registering these seed locations with the corresponding locations in the previously acquired treatment planning computed tomographic (CT) scan, it is possible to adjust the patient position so that seed displacement is accommodated. The authors present an unsupervised automatic algorithm that identifies seeds in both planning and pretreatment images and subsequently determines a rigid geometric transformation between the two sets. The algorithm is applied to the imaging series of ten prostate cancer patients. Each test series is comprised of a single multislice planning CT and multiple megavoltage conebeam (MVCB) images. Each MVCB dataset is obtained immediately prior to a subsequent treatment session. Seed locations were determined to within 1 mm with an accuracy of 97 +/- 6.1% for datasets obtained by application of a mean imaging dose of 3.5 cGy per study. False positives occurred in three separate instances, but only when datasets were obtained at imaging doses too low to enable fiducial resolution by a human operator, or when the prostate gland had undergone large displacement or significant deformation. The registration procedure requires under nine seconds of computation time on a typical contemporary computer workstation.

Focused beam-stop array for the measurement of scatter in megavoltage portal and cone beam CT imaging.

We describe a focused beam-stop array (BSA) for the measurement of object scatter in imaging systems that utilize x-ray beams in the megavoltage (MV) energy range. The BSA consists of 64 doubly truncated tungsten cone elements of 0.5 cm maximum diameter that are arranged in a regular array on an acrylic slab. The BSA is placed in the accessory tray of a medical linear accelerator at a distance of approximately 50 cm from the focal spot. We derive an expression that allows us to estimate the scatter in an image taken without the array present, given image values in a second image with the array in place. The presence of the array reduces fluence incident on the imaged object. This leads to an object-dependent underestimation bias in the scatter measurements. We apply corrections in order to address this issue. We compare estimates of the flat panel detector response to scatter obtained using the BSA to those derived from Monte Carlo simulations. We find that the two estimates agree to within 10% in terms of RMS error for 30 cm x 30 cm water slabs in the thickness range of 10-30 cm. Larger errors in the scatter estimates are encountered for thinner objects, probably owing to extrafocal radiation sources. However, RMS errors in the estimates of primary images are no more than 5% for water slab thicknesses in the range of 1-30 cm. The BSA scatter estimates are also used to correct cone beam tomographic projections. Maximum deviations of central profiles of uniform water phantoms are reduced from 193 to 19 HU after application of corrections for scatter, beam hardening, and lateral truncation that are based on the BSA-derived scatter estimate. The same corrections remove the typical cupping artifact from both phantom and patient images. The BSA proves to be a useful tool for quantifying and removing image scatter, as well as for validating models of MV imaging systems.

CT truncation artifact removal using water-equivalent thicknesses derived from truncated projection data.

PURPOSE: Large patient anatomies and limited imaging field-of-view (FOV) lead to truncation of CT projections. Truncation introduces serious artifacts into reconstructed images, including central cupping and bright external rings. FOV may be increased using laterally offset detectors, but this requires sophisticated imaging hardware and full angular scanning. We propose a novel method to complete truncated projections based on the observation that the thickness of the patient may be estimated along the projection rays by calculating water-equivalent thicknesses (WET). These values are not at all affected by truncation and thus constitute valuable auxiliary information. METHODS: We parameterize pairs of points along each ray that intersects the unknown object boundary. These points are separated by the measured WET value (obtained from projections that have been corrected for scatter and beam-hardening). We assume, for all large body parts, that the patient outline may be roughly approximated as an ellipse. Using a deterministic optimization algorithm, we simultaneously estimate the point positions and ellipse parameters by minimizing the distance between point sets and the ellipse boundary. The optimal ellipse is used to complete the truncated projections. Reconstruction then ensues. We apply the algorithm to a severely truncated CT dataset of a typical abdomen. RESULTS: The RMS error between complete data and truncated reconstructions (corrected using an empirical extrapolation approach) is 20.4% for an abdominal dataset. The new algorithm reduces this error to 1.0%. CONCLUSION: Even thought the algorithm assumes an elliptical patient cross-section, truly impressive increases in quantitative image quality are observed. The presence of pelvic bone in the image does not appreciably bias the ellipse position even though it does bias the thickness estimates for some rays. The algorithm incurs low computational cost and is suitable for on-line clinical workflows.

Dose calculation using megavoltage cone-beam CT.

PURPOSE: To demonstrate the feasibility of performing dose calculation on megavoltage cone-beam CT (MVCBCT) of head-and-neck patients in order to track the dosimetric errors produced by anatomic changes. METHODS AND MATERIALS: A simple geometric model was developed using a head-size water cylinder to correct an observed cupping artifact occurring with MVCBCT. The uniformity-corrected MVCBCT was calibrated for physical density. Beam arrangements and weights from the initial treatment plans defined using the conventional CT were applied to the MVCBCT image, and the dose distribution was recalculated. The dosimetric inaccuracies caused by the cupping artifact were evaluated on the water phantom images. An ideal test patient with no observable anatomic changes and a patient imaged with both CT and MVCBCT before and after considerable weight loss were used to clinically validate MVCBCT for dose calculation and to determine the dosimetric impact of large anatomic changes. RESULTS: The nonuniformity of a head-size water phantom ( approximately 30%) causes a dosimetric error of less than 5%. The uniformity correction method developed greatly reduces the cupping artifact, resulting in dosimetric inaccuracies of less than 1%. For the clinical cases, the agreement between the dose distributions calculated using MVCBCT and CT was better than 3% and 3 mm where all tissue was encompassed within the MVCBCT. Dose-volume histograms from the dose calculations on CT and MVCBCT were in excellent agreement. CONCLUSION: MVCBCT can be used to estimate the dosimetric impact of changing anatomy on several structures in the head-and-neck region.