Identifying the optimal deep learning architecture and parameters for automatic beam aperture definition in 3D radiotherapy

PURPOSE

Two-dimensional radiotherapy is often used to treat cervical cancer in low- and middle-income countries, but treatment planning can be challenging and time-consuming. Neural networks offer the potential to greatly decrease planning time through automation, but the impact of the wide range of hyperparameters to be set during training on model accuracy has not been exhaustively investigated. In the current study, we evaluated the effect of several convolutional neural network architectures and hyperparameters on 2D radiotherapy treatment field delineation.

METHODS

Six commonly used deep learning architectures were trained to delineate four-field box apertures on digitally reconstructed radiographs for cervical cancer radiotherapy. A comprehensive search of optimal hyperparameters for all models was conducted by varying the initial learning rate, image normalization methods, and (when appropriate) convolutional kernel size, the number of learnable parameters via network depth and the number of feature maps per convolution, and nonlinear activation functions. This yielded over 1700 unique models, which were all trained until performance converged and then tested on a separate dataset.

RESULTS

Of all hyperparameters, the choice of initial learning rate was most consistently significant for improved performance on the test set, with all top-performing models using learning rates of 0.0001. The optimal image normalization was not consistent across architectures. High overlap (mean Dice similarity coefficient = 0.98) and surface distance agreement (mean surface distance < 2 mm) were achieved between the treatment field apertures for all architectures using the identified best hyperparameters. Overlap Dice similarity coefficient (DSC) and distance metrics (mean surface distance and Hausdorff distance) indicated that DeepLabv3+ and D-LinkNet architectures were least sensitive to initial hyperparameter selection.

CONCLUSION

DeepLabv3+ and D-LinkNet are most robust to initial hyperparameter selection. Learning rate, nonlinear activation function, and kernel size are also important hyperparameters for improving performance.

Intrafractional Diaphragm Variations During Breath-Hold Stereotactic Body Radiotherapy for a Liver Tumor Based on Real-Time Registration Between Kilovoltage Projection Streaming and Digitally Reconstructed Radiograph Images: A Case Report

In liver stereotactic body radiotherapy (SBRT), precise image guidance is paramount, serving as the foundation of this treatment approach. The accuracy of SBRT in liver cancer treatment heavily relies on meticulous imaging techniques. The diaphragm, situated adjacent to the liver, is a crucial anatomical structure susceptible to positional and motion variations, which can potentially impact the accuracy of liver tumor targeting. This study explores the application of real-time kilovoltage projection streaming images (KVPSI) in comparison to digitally reconstructed radiography (DRR) for assessing diaphragm position deviations during breath-hold liver tumor SBRT. A 76-year-old male diagnosed with cholangiocarcinoma underwent breath-hold SBRT using split arc volumetric modulated arc therapy (VMAT), where a full arc was split into six sub-arcs, each spanning 60 degrees. The diaphragm dome positions were continuously monitored through KVPSI during treatment. The intrafractional position deviations of the diaphragm were calculated and analyzed for each split arc. The case report revealed a mean diaphragm dome deviation of 0.47 mm (standard deviation: 4.47 mm) in the entire arc. This pioneering study showcases the feasibility of intrafractional diaphragm position variation assessment using real-time KVPSI during the breath-hold liver tumor VMAT-SBRT. Integrating real-time imaging techniques enhances our comprehension of the intra-breath-hold variations, thereby guiding adaptive treatment strategies and potentially improving treatment outcomes. Clinical validation through further research is essential.

Machine learning for automated and real-time two-dimensional to three-dimensional registration of the spine using a single radiograph

OBJECTIVE

The goal of this work was to methodically evaluate, optimize, and validate a self-supervised machine learning algorithm capable of real-time automatic registration and fluoroscopic localization of the spine using a single radiograph or fluoroscopic frame.

METHODS

The authors propose a two-dimensional to three-dimensional (2D-3D) registration algorithm that maximizes an image similarity metric between radiographic images to identify the position of a C-arm relative to a 3D volume. This work utilizes digitally reconstructed radiographs (DRRs), which are synthetic radiographic images generated by simulating the x-ray projections as they would pass through a CT volume. To evaluate the algorithm, the authors used cone-beam CT data for 127 patients obtained from an open-source de-identified registry of cervical, thoracic, and lumbar scans. They systematically evaluated and tuned the algorithm, then quantified the convergence rate of the model by simulating C-arm registrations with 80 randomly simulated DRRs for each CT volume. The endpoints of this study were time to convergence, accuracy of convergence for each of the C-arm's degrees of freedom, and overall registration accuracy based on a voxel-by-voxel measurement.

RESULTS

A total of 10,160 unique radiographic images were simulated from 127 CT scans. The algorithm successfully converged to the correct solution 82% of the time with an average of 1.96 seconds of computation. The radiographic images for which the algorithm converged to the solution demonstrated 99.9% registration accuracy despite utilizing only single-precision computation for speed. The algorithm was found to be optimized for convergence when the search space was limited to a ± 45° offset in the right anterior oblique/left anterior oblique, cranial/caudal, and receiver rotation angles with the radiographic isocenter contained within 8000 cm3 of the volumetric center of the CT volume.

CONCLUSIONS

The investigated machine learning algorithm has the potential to aid surgeons in level localization, surgical planning, and intraoperative navigation through a completely automated 2D-3D registration process. Future work will focus on algorithmic optimizations to improve the convergence rate and speed profile.

Assessment of Flatfoot Deformity Using Digitally Reconstructed Radiographs: Reliability and Comparison to Conventional Radiographs

BACKGROUND

Digitally reconstructed radiographs (DRRs) generated from weightbearing computed tomography (WBCT) may potentially substitute for weightbearing plain radiographs (XRs) but have not been clinically validated. This study aims to test the reliability of 6 radiographic parameters of progressive collapsing foot deformity (PCFD) as measured on DRR, to investigate whether DRR represents comparably to XR through the same measurements, and to compare agreement of DRR and XR measurements of a standardized arch height parameter with reference measurements made on WBCT.

METHODS

DRR generated from preoperative WBCT of 71 patients (72 feet) treated surgically for PCFD were retrospectively compared with preoperative weight-bearing XR after exclusion criteria were applied. Six radiographic measurements were performed, including Meary angle, calcaneal pitch (CPA), medial cuneiform height (MCH), AP talar-first metatarsal angle (T-1MT), talonavicular coverage (TNCA), and talar incongruency (TIA). Arch height was measured on XR, DRR, and WBCT using a validated, standardized, navicular-based index. Intraclass correlation coefficients assessed DRR intraobserver and interobserver reliability. Paired samples t tests tested differences between XR and DRR. Bland-Altman limits of agreement analysis compared DRR and XR agreement with WBCT measurements.

RESULTS

Measurements were within standard PCFD ranges on XR and DRR. All measurements demonstrated excellent intrarater reliability and good to excellent interrater agreement, consistent with previous literature on XR. No differences were found for Meary, CPA, or TNCA. Minor differences were observed for MCH, T-1MT, and TIA. DRR measurements demonstrated greater agreement with WBCT than XR measurements.

CONCLUSION

DRR from WBCT may be a promising substitute for XR in the clinical evaluation of PCFD. Radiographic measurements made on DRR demonstrated good to excellent reliability. Although small differences were found between XR and DRR for certain measurements, DRR more accurately represented medial arch anatomy compared to gold standard WBCT data than XR. If validated as a clinical substitute, DRR could eventually obviate XR where WBCT is available.

LEVEL OF EVIDENCE

Level III, retrospective cohort study.

Differences in image density adjustment parameters on the image matching accuracy of a floor-mounted kV X-ray image-guided radiation therapy system

This study aimed to investigate the effect of two different image density adjustment parameters on the results of image matching at six degrees of freedom using radiographic images generated by the ExacTrac X‐ray system in brain stereotactic radiosurgery (SRS). This study comprised 32 patients who underwent brain SRS at our hospital from January 2020 to December 2020. In this study, (1) the default parameter (an image density parameter between “tissue” and “bone”) was an image density parameter for digitally reconstructed radiograph (DRR) generation used at many facilities, and (2) the bone parameter was the steepest contrast parameter used at our hospital. Of the 32 patients, 24 (75%) had a couch angle of 0.5 mm or more in the translational direction or 0.5° or more in the rotational direction, and 10 (31%) had a couch angle of 1.0 mm or more in the translational direction or 1.0° or more in the rotational direction. Among the 131 cases of all couch angles, 46 (35%) cases had a translational direction of 0.5 mm or more or a rotational direction of 0.5° or more, and 15 (11%) had a translational direction of 1.0 mm or more or a rotational direction of 1.0° or more. The results of this study indicate the usefulness of using appropriate DRR parameters for each case, rather than using the default settings. The use of appropriate DRR parameters can lead to accurate position matching results, leading to fewer image‐guided radiation therapy shots and a lower imaging dose.

A comparison of proton stopping power measured with proton CT and x-ray CT in fresh postmortem porcine structures

PURPOSE

Currently, calculations of proton range in proton therapy patients are based on a conversion of CT Hounsfield units of patient tissues into proton relative stopping power. Uncertainties in this conversion necessitate larger proximal and distal planned target volume margins. Proton CT can potentially reduce these uncertainties by directly measuring proton stopping power. We aim to demonstrate proton CT imaging with complex porcine samples, to analyze in detail three-dimensional regions of interest, and to compare proton stopping powers directly measured by proton CT to those determined from x-ray CT scans.

METHODS

We have used a prototype proton imaging system with single proton tracking to acquire proton radiography and proton CT images of a sample of porcine pectoral girdle and ribs, and a pig's head. We also acquired close in time x-ray CT scans of the same samples and compared proton stopping power measurements from the two modalities. In the case of the pig's head, we obtained x-ray CT scans from two different scanners and compared results from high-dose and low-dose settings.

RESULTS

Comparing our reconstructed proton CT images with images derived from x-ray CT scans, we find agreement within 1% to 2% for soft tissues and discrepancies of up to 6% for compact bone. We also observed large discrepancies, up to 40%, for cavitated regions with mixed content of air, soft tissue, and bone, such as sinus cavities or tympanic bullae.

CONCLUSIONS

Our images and findings from a clinically realistic proton CT scanner demonstrate the potential for proton CT to be used for low-dose treatment planning with reduced margins.

Automatic generation of three-dimensional dose reconstruction data for two-dimensional radiotherapy plans for historically treated patients

Performing large-scale three-dimensional radiation dose reconstruction for patients requires a large amount of manual work. We present an image processing-based pipeline to automatically reconstruct radiation dose. The pipeline was designed for childhood cancer survivors that received abdominal radiotherapy with anterior-to-posterior and posterior-to-anterior field set-up. First, anatomical landmarks are automatically identified on two-dimensional radiographs. Second, these landmarks are used to derive parameters to emulate the geometry of the plan on a surrogate computed tomography. Finally, the plan is emulated and used as input for dose calculation. For qualitative evaluation, 100 cases of automatic and manual plan emulations were assessed by two experienced radiation dosimetrists in a blinded comparison. The two radiation dosimetrists approved 100%/100% and 92%/91% of the automatic/manual plan emulations, respectively. Similar approval rates of 100% and 94% hold when the automatic pipeline is applied on another 50 cases. Further, quantitative comparisons resulted in on average < 5    mm difference in plan isocenter/borders, and < 0.9    Gy in organ mean dose (prescribed dose: 14.4 Gy) calculated from the automatic and manual plan emulations. No statistically significant difference in terms of dose reconstruction accuracy was found for most organs at risk. Ultimately, our automatic pipeline results are of sufficient quality to enable effortless scaling of dose reconstruction data generation.

[Evaluation of the Image Registration Program for Portal Images Using CR and DRR Images in Radiation Therapy]

In this study, computer simulations and experiments were used to verify the accuracy of a two-dimensional image registration program (program) for portal images that we previously developed. The program used a computed radiography cassette system and digitally reconstructed radiography images as planning images for external beam radiation therapy. Using this program, we also investigated the reason two-dimensional automatic image registration images experienced large misregistration in clinical practice using commercial image registration systems. Mutual information and normalized mutual information were used as the registration criteria. To investigate the influence of image background with or without a region of interest (ROI), results of image registrations were compared. Parameters of image registration were defined as translation in the horizontal and vertical directions (x and y, respectively) and rotation (θ) around the axis perpendicular to the x-y plane. There was no significant difference in image registration arising from the difference between mutual information and normalized mutual information. Image registration was improved with a ROI. Regardless of the registration criteria, errors in image registration with a ROI in the experimental study were ≤1.2 mm in directions x and y and ≤1.0 degree in rotation θ. We found that image registration required setting up as close to the planned position as possible.

The critical acromial point: the anatomic location of the lateral acromion in the critical shoulder angle

BACKGROUND

Acromioplasty has been proposed as a means of altering elevated critical shoulder angles (CSAs). We aimed to localize the critical acromion point (CAP) responsible for the acromial contribution of the CSA and determine whether resection of the CAP can alter the CSA to a normal range.

METHODS

The CAP and 3-dimensional (3D) CSAs were determined on 3D computed tomography reconstructions of 88 cadaveric shoulders and compared with corresponding CSAs on digitally reconstructed radiographs. The position of the CAP was fluoroscopically isolated in 20 of these specimens and the resulting fluoroscopic CSA compared with the corresponding 3D CAP and 3D CSA. We investigated the CSA before and after a virtual acromioplasty of 2.5 and 5 mm at the CAP in specimens with a CSA greater than 35°.

RESULTS

The mean CAP was 21% ± 10% of the acromial anterior-posterior length from the anterolateral corner. There was no difference between the mean 3D CSA and the CSA on digitally reconstructed radiographs (32° vs 32°, P = .096). No difference between the mean fluoroscopic CSA and 3D CSA was found (31° vs 31°, P = .296). A 2.5-mm acromial resection failed to reduce the CSA to 35° or less in 7 of 13 shoulders, whereas a 5-mm resection reduced the CSA to 35° or less in 12 of 13.

CONCLUSION

The CAP was localized to the anterolateral acromial edge and was easily identified fluoroscopically. A 5-mm acromial resection was effective in reducing the CSA to 35° or less. These data can guide surgeons in where and how to alter the CSA if future studies demonstrate a clinical benefit to surgically modifying this radiographic parameter.

The influence of radiographic viewing perspective and demographics on the critical shoulder angle

BACKGROUND

Accurate assessment of the critical shoulder angle (CSA) is important in clinical evaluation of degenerative rotator cuff tears. This study analyzed the influence of radiographic viewing perspective on the CSA, developed a classification system to identify malpositioned radiographs, and assessed the relationship between the CSA and demographic factors.

METHODS

Glenoid height, width, and retroversion were measured on 3-dimensional computed tomography reconstructions of 68 cadaver scapulae. A digitally reconstructed radiograph was aligned perpendicular to the scapular plane, and retroversion was corrected to obtain a true anteroposterior (AP) view. In 10 scapulae, incremental anteversion/retroversion and flexion/extension views were generated. The CSA was measured, and a clinically applicable classification system was developed to detect views with >2° change in CSA vs. true AP view.

RESULTS

The average CSA was 33° ± 4°. Intraobserver and interobserver reliability was high (intraclass correlation coefficient ≥ 0.81) but decreased with increasing viewing angle. Views beyond 5° anteversion, 8° retroversion, 15° flexion, and 26° extension resulted in >2° deviation of the CSA compared with the true AP view. The classification system was capable of detecting aberrant viewing perspectives with sensitivity of 95% and specificity of 53%. Correlations between glenoid size and CSA were small (R ≤ 0.3), and CSA did not vary by gender (P = .426) or side (P = .821).

CONCLUSIONS

The CSA was most susceptible to malposition in anteversion/retroversion. Deviations as little as 5° in anteversion resulted in a CSA >2° from true AP view. A new classification system refines the ability to collect true AP radiographs of the scapula. The CSA was unaffected by demographic factors.