Ground Truth Contours Research Articles

Real-Time Motion Monitoring Using Orthogonal Cine MRI during MR-Guided Radiation Therapy for Liver Cancer

<h3>Purpose/Objective(s)</h3> This work investigates the feasibility of real-time motion monitoring (MM) using orthogonal cine MRI during MR guided Adaptative Radiation Therapy (MRgART) for liver tumors that are normally invisible on cine MRI with no contrast agent. <h3>Materials/Methods</h3> An investigational package, Motion Monitoring Research Package (MMRP), was developed to measure real-time target motion between orthogonal cine MRI and the daily baseline 3D MRI. The MMRP package was evaluated on 8 liver cancer patients treated on a 1.5T MR-Linac. For all patients, a 3D mid-position (MidP) image derived from a daily 4D MRI was used to delineate a sub-liver volume (SLV) encompassing the tumor. During the same session, real-time 2D T2/T1-weighted bFFE cine MRI were collected with free breathing during MRgART. The cine images were acquired with a temporal resolution of 200 ms interleaved between coronal and sagittal orientations. MMRP workflow included two steps: (1) creating a 2D template in each orientation using cine data acquired over the first 12 s, followed by a rigid registration between the combined templates and the MidP image over the SLV region; (2) measuring the SLV motion by registering the rest of the 2D cine series with the corresponding template. To evaluate the MM accuracy from MMRP, the ground-truth contours on cine MRI were created manually. Common visible vessels and segments of liver boundaries in proximity to the tumor were used as anatomical landmarks for reproducible delineations on both the 3D and cine images. MM accuracy was measured using the standard deviation of the error (SDE) to compare the center of mass position of the ground-truth and the monitored motion. Maximum of tumor motion was assessed on 4D MRI. <h3>Results</h3> In 6 of the 8 cases, the assessed mean (range) centroid motions were 8.8 (4.6-13.8), 1.5 (0.7-2.8) and 3.2 (1.5-5) mm with mean SDE (range) values of 1.2 (0.7-1.6), 0.94 (0.4-1.7), and 1 (0.4-1.4) mm in superior inferior (SI), left right, and anterior posterior (AP) directions, respectively. The mean (range) of the maximum tumor motion from the 4D MRI was 8 (5-11) mm in SI direction, smaller than the centroid MM, indicating the importance of motion capture using the MM in real-time. For the remaining two cases, quantitative MM was challenging due to target deformation and large through-plane motion in the AP direction. <h3>Conclusion</h3> We have demonstrated the feasibility of accurate real-time MM of a liver tumor using a sub-liver volume as a tumor surrogate, based on real-time orthogonal cine MRI during MRgART. The approach overcomes difficulty due to the invisibility of liver tumor on the T2/T1-w 2D cine without using a contrast agent and the need of implanting radio-opaque clips or the approximation of the whole liver motion as the tumor motion. With further investigation with more patient data, the real-time MM technique may be implemented in MRgART to manage intrafraction liver motion.

Stratified Assessment of a Commercial Deep Learning Algorithm for Automated Detection and Contouring of Metastatic Brain Tumors in Stereotactic Radiosurgery

<h3>Purpose/Objective(s)</h3> Artificial intelligence-based tools can be leveraged to improve detection and segmentation of brain metastases for stereotactic radiosurgery (SRS). This study explores a deep learning algorithm with recent FDA clearance to assist in brain tumor contouring. In this study, we aimed to assess the performance of this tool by various demographic and clinical characteristics among patients with brain metastases treated with SRS. <h3>Materials/Methods</h3> We randomly selected 100 patients with brain metastases who underwent initial SRS on a frameless robotic radiosurgery system from 2017 to 2020 at a single institution. Cases with resection cavities were excluded from the analysis. Computed tomography (CT) and axial T1-weighted post-contrast magnetic resonance (MR) image data were extracted for each patient and uploaded to VBrain. A brain metastasis was considered "detected" when the VBrain "predicted" contours overlapped with the corresponding physician contours ("ground-truth" contours). We evaluated performance of VBrain against ground-truth contours using the following metrics: lesion-wise Dice similarity coefficient (DSC), lesion-wise average Hausdorff distance (AVD), false positive count (FP), and lesion-wise sensitivity (%). Wilcoxon rank-sum tests were performed to assess the relationships between patient characteristics including sex, race, primary histology, age, and size and number of brain metastases, and performance metrics such as DSC, AVD, FP, and sensitivity. <h3>Results</h3> We analyzed 100 patients with 422 intact brain metastases treated with SRS. Our cohort consisted of patients with a median number of 2 brain metastases (range: 1 to 47), median age of 79 (range: 19 to 91), and 50% male and 50% female patients. The primary site breakdown was 56% lung, 10% melanoma, 9% breast, 8% gynecological, 5% gastrointestinal, 2% prostate, and 5% other, while the race breakdown was 60% White, 18% Asian, 3% Black/African American, 2% Native Hawaiian or other Pacific Islander, and 17% other/unknown/not reported. The median tumor size was 0.118 c.c. (range: 0.003-28.838 c.c.). We found mean lesion-wise DSC to be 0.714, mean lesion-wise AVD to be 8.44% of lesion size (0.796 mm), mean FP count to be 0.17 tumors per case, and lesion-wise sensitivity to be 83.65% for all lesions. Moreover, mean sensitivity was found to be 98%, 91.93%, and 79.87% for lesions with diameter equal to and greater than 10 mm, 5 mm, and 3 mm, respectively. No other significant differences in performance metrics were observed across demographic or clinical characteristic groups. <h3>Conclusion</h3> A commercial deep learning algorithm showed promising results in segmenting brain metastases, with 98% sensitivity for metastases with diameters of 1 cm or higher. Performance declined for smaller lesions. As advances in imaging and radiosurgery enable the detection and treatment of smaller brain metastases, future work is ongoing to develop and improve these tools to assist in radiosurgical treatment planning.

Automated Brain Metastasis Detection and Segmentation Using Deep-Learning Method

<h3>Purpose/Objective(s)</h3> Brain metastases (BM) detection and segmentation can be difficult and time-consuming during stereotactic radiosurgery (SRS) planning. This work aims to develop deep learning models to automatically detect and segment BM on MRI. <h3>Materials/Methods</h3> Two volumetric deep convolutional neural network (V-net) models, general (GM) and small lesion models (SM), were developed using T1-weighted gadolinium contrast enhanced and T2-weighted FLAIR MRIs along with ground truth contours of all BMs collected for 53 patients from our own institution. The GM was developed by transfer learning based on a model previously trained with multi-institutional datasets. The SM was a new model presently trained using images slices with small lesions (volumes < 1 cc). The training sets for both models were the data from 43 patients, a total of 5,204 image sets extracted with three different patch sizes (16 × 16 × 16, 32 × 32 × 32, and 64 × 64 × 64). The obtained models were evaluated using a test set of 10 patients in terms of the Dice similarity coefficient (DSC), sensitivity, specificity, and area under the receiver operating characteristic curve (AUC). <h3>Results</h3> For the GM, the voxel-level segmentation AUCs were 0.90 +/- 0.10, 0.88 +/- 0.10, and 0.87 +/- 0.11 for patch sizes of 16 × 16 × 16, 32 × 32 × 32, and 64 × 64 × 64, respectively, and the detection sensitivity and specificity were 0.86 and 0.94, respectively, for all test datasets (all lesions). For the test data excluding small lesions, the sensitivity and specificity became 0.93 and 0.95, respectively. For SM, a sensitivity of 0.76 and a specificity of 0.91 were achieved for all small lesions in the test datasets. The DSC was 0.92 for GM for all lesions and was 0.82 for SM for all the small BMs. <h3>Conclusion</h3> The two deep learning models, SM and GM, can automatically detect and segment brain metastases with reasonable accuracy, sensitivity, and specificity on the MRIs commonly used for SRS planning. For BMs with volumes > 1 cc, the GM achieved clinically relevant results. With further development using larger datasets, the models can be implemented to improve the clinical practice of SRS planning.

MRI-based prostate and dominant lesion segmentation using cascaded scoring convolutional neural network.

Dose escalation to dominant intraprostatic lesions (DILs) is a novel treatment strategy to improve the treatment outcome of prostate radiation therapy. Treatment planning requires accurate and fast delineation of the prostate and DILs. In this study, a 3D cascaded scoring convolutional neural network is proposed to automatically segment the prostate and DILs from MRI. The proposed cascaded scoring convolutional neural network performs end-to-end segmentation by locating a region-of-interest (ROI), identifying the object within the ROI, and defining the target. A scoring strategy, which is learned to judge the segmentation quality of DIL, is integrated into cascaded convolutional neural network to solve the challenge of segmenting the irregular shapes of the DIL. To evaluate the proposed method, 77 patients who underwent MRI and PET/CT were retrospectively investigated. The prostate and DIL ground truth contours were delineated by experienced radiologists. The proposed method was evaluated with fivefold cross-validation and holdout testing. The average centroid distance, volume difference, and Dice similarity coefficient (DSC) value for prostate/DIL are 4.3±7.5/3.73±3.78mm, 4.5±7.9/0.41±0.59cc, and 89.6±8.9/84.3±11.9%, respectively. Comparable results were obtained in the holdout test. Similar or superior segmentation outcomes were seen when compared the results of the proposed method to those of competing segmentation approaches. The proposed automatic segmentation method can accurately and simultaneously segment both the prostate and DILs. The intended future use for this algorithm is focal boost prostate radiation therapy.

Automatic Contour Refinement for Deep Learning Auto-segmentation of Complex Organs in MRI-guided Adaptive Radiation Therapy

PurposeFast and accurate auto-segmentation on daily images is essential for magnetic resonance imaging (MRI)–guided adaptive radiation therapy (ART). However, the state-of-the-art auto-segmentation based on deep learning still has limited success, particularly for complex structures in the abdomen. This study aimed to develop an automatic contour refinement (ACR) process to quickly correct for unacceptable auto-segmented contours. Methods and MaterialsAn improved level set–based active contour model (ACM) was implemented for the ACR process and was tested on the deep learning–based auto-segmentation of 80 abdominal MRI sets along with their ground truth contours. The performance of the ACR process was evaluated using 4 contour accuracy metrics: the Dice similarity coefficient (DSC), mean distance to agreement (MDA), surface DSC, and added path length (APL) on the auto-segmented contours of the small bowel, large bowel, combined bowels, pancreas, duodenum, and stomach. ResultsA portion (3%-39%) of the corrected contours became practically acceptable per the American Association of Physicists in Medicine Task Group 132 (TG-132) recommendation (DSC >0.8 and MDA <3 mm). The best correction performance was seen in the combined bowels, where for the contours with major errors (initial DSC <0.5 or MDA >8 mm), the mean DSC increased from 0.34 to 0.59, the mean MDA decreased from 7.02 mm to 5.23 mm, and the APL reduced by almost 20 mm, whereas for the contours with minor errors, the mean DSC increased from 0.72 to 0.79, the mean MDA decreased from 3.35 mm to 3.29 mm, and more than one-third (39%) of the ACR contours became clinically acceptable. The execution time for the ACR process on one subregion was less than 2 seconds using an NVIDIA GTX 1060 GPU. ConclusionsThe ACR process implemented based on the ACM was able to quickly correct for some inaccurate contours produced from MRI-based deep learning auto-segmentation of complex abdominal anatomy. The ACR method may be integrated into the auto-segmentation step to accelerate the process of MRI-guided ART.

Automatic Delineation of Gross Tumor Volume Based on Magnetic Resonance Imaging by Performing a Novel Semisupervised Learning Framework in Nasopharyngeal Carcinoma

We aimed to validate the accuracy and clinical value of a novel semisupervised learning framework for gross tumor volume (GTV) delineation in nasopharyngeal carcinoma. Two hundred fifty-eight patients with magnetic resonance imaging data sets were divided into training (n = 180), validation (n = 20), and testing (n = 58) cohorts. Ground truth contours of nasopharynx GTV (GTVnx) and node GTV (GTVnd) were manually delineated by 2 experienced radiation oncologists. Twenty percent (n = 36) labeled and 80% (n = 144) unlabeled images were used to train the model, producing model-generated contours for patients from the testing cohort. Nine experienced experts were invited to revise model-generated GTV in 20 randomly selected patients from the testing cohort. Six junior oncologists were asked to delineate GTV in 12 randomly selected patients from the testing cohort without and with the assistance of the model, and revision degrees were compared under these 2 modes. The Dice similarity coefficient (DSC) was used to quantify the accuracy of the model. The model-generated contours showed a high accuracy compared with ground truth contours, with an average DSC score of 0.83 and 0.80 for GTVnx and GTVnd, respectively. There was no significant difference in DSC score between T1-2 and T3-4 patients (0.81 vs 0.83; P = .223), or between N1-2 and N3 patients (0.80 vs 0.79; P = .807). The mean revision degree was lower than 10% in 19 (95%) patients for GTVnx and in 16 (80%) patients for GTVnd. With assistance of the model, the mean revision degree for GTVnx and GTVnd by junior oncologists was reduced from 25.63% to 7.75% and from 21.38% to 14.44%, respectively. Meanwhile, the delineating efficiency was improved by over 60%. The proposed semisupervised learning-based model showed a high accuracy for delineating GTV of nasopharyngeal carcinoma. It was clinically applicable and could assist junior oncologists to improve GTV contouring accuracy and save contouring time.

Fully automated lumen and vessel contour segmentation in intravascular ultrasound datasets.

Segmentation of lumen and vessel contours in intravascular ultrasound (IVUS) pullbacks is an arduous and time-consuming task, which demands adequately trained human resources. In the present study, we propose a machine learning approach to automatically extract lumen and vessel boundaries from IVUS datasets. The proposed approach relies on the concatenation of a deep neural network to deliver a preliminary segmentation, followed by a Gaussian process (GP) regressor to construct the final lumen and vessel contours. A multi-frame convolutional neural network (MFCNN) exploits adjacency information present in longitudinally neighboring IVUS frames, while the GP regression method filters high-dimensional noise, delivering a consistent representation of the contours. Overall, 160 IVUS pullbacks (63 patients) from the IBIS-4 study (Integrated Biomarkers and Imaging Study-4, Trial NCT00962416), were used in the present work. The MFCNN algorithm was trained with 100 IVUS pullbacks (8427 manually segmented frames), was validated with 30 IVUS pullbacks (2583 manually segmented frames) and was blindly tested with 30 IVUS pullbacks (2425 manually segmented frames). Image and contour metrics were used to characterize model performance by comparing ground truth (GT) and machine learning (ML) contours. Median values (interquartile range, IQR) of the Jaccard index for lumen and vessel were 0.913, [0.882,0.935] and 0.940, [0.917,0.957], respectively. Median values (IQR) of the Hausdorff distance for lumen and vessel were 0.196mm, [0.146,0.275]mm and 0.163mm, [0.122,0.234]mm, respectively. Also, the mean value of lumen area predictions, and limits of agreement were -0.19mm2, [1.1,-1.5]mm2, while the mean value and limits of agreement of plaque burden were 0.0022, [0.082,-0.078]. The results obtained with the model developed in this work allow us to conclude that the proposed machine learning approach delivers accurate segmentations in terms of image metrics, contour metrics and clinically relevant variables, enabling its use in clinical routine by mitigating the costs involved in the manual management of IVUS datasets.

A Double-Blind Study to Evaluate the Feasibility of Using AI-Powered Auto-Segmentation in Prostate Cancer Treatment

To evaluate whether the contours generated by an AI-powered auto-segmentation software can be used directly for prostate cancer radiation therapy.A total of 23 prostate cases treated between 2018 and 2020 were retrospectively randomly selected. The prostate and surrounding organs at risk (OARs) were delineated on computed tomography (CT) scans by a radiation oncologist (RO) as ground-truth (GT) contour and by Artificial Intelligence (AI) model as AI contour. The GT and AI contour sets were anonymized randomly to be evaluated by a different RO. To eliminate the planner's bias, RapidPlans were used to generate treatment plans using AI and GT contours. The OAR dose on both plans were evaluated on GT contour sets to determine whether the AI contour can produce similar results as the GT contour.In subjective double-blind test, AI contours show that 95.7% (22 out of 23) are scored as either "Great" (34.8%) or "Acceptable without changes" (60.9%), and only one case (4.3%) is evaluated as "Acceptable with minor changes". Note that the GT contours also have one case in the "minor change" category. Totally 69.6% (16 of 23) AI contours were considered equal or better than the GT counterparts. Paired t-tests showed no significant differences in dosimetric plan quality yielded by the AI and GT for the prostate, rectum, femoral heads, and penile bulb (P > 0.05). Statistically significant difference (P < 0.05) was observed on Bladder and Seminal Vesicles. However, no plan fails the RTOG-0815 organ tolerances.The investigated auto-segmentation model for prostate anatomy provides compatible performance to manually delineated contours by a Radiation Oncologist.

Impact of MRI Sequence Resolution for Target Volume Definition in Stereotactic Radiosurgery

MRI is a standard image modality in cranial radiosurgery (SRS) for defining the target volume. There is disparity with different MRI sequences however due to differences in acquisition resolution, which can directly impact accuracy of target segmentation. Here we acquired MRI using different sequences on phantoms to evaluate the effect on volume delineation in the context of cranial SRS.Four different T1-weight MR pulse sequences were included: (1) SE5mm: axial and coronal Spin Echo (SE) 2D acquisition with 0.4 × 0.4 mm2 in-plane resolution and 5 mm cross-plane; (1) SE3mm: axial and coronal SE with 0.5 × 0.5 mm2 in-plane and 3 mm cross-plane; (3) TFE: gradient echo 3D acquisition with 0.92 × 0.92 mm2 in axial plane and 1.25 mm cross-plane; (4) BRAVO: gradient echo 3D with 0.4 × 0.4 mm2 in axial plane and 0.5 mm cross-plane. Four phantoms with different shape and volume (0.54 to 25 cm3) were imaged, resulting in 16 sets of MRIs. Four radiation oncologists provided contours on individual sets of MRIs. All observer contours were compared with ground truth, which was defined on CT image according to the absolute dimensions of the phantom structures. Dice coefficients (DSC), Hausdorff distance (MaxHD, MeanHD, σHD) as well as ratio between reconstructed and true volume (Ratiovol) were evaluated between observer and ground truth contours. A two-sided signed-rank test was performed to determine whether differences were statistically significant.The comparison to ground truth as a function of MR sequence is presented in the Table. All indices improved as the MR resolution increased from SE5mm to BRAVO. All differences, with the exception of Ratiovol between SE3mm and TFE, were statistically significant (P < 0.01). Inter-observer variation which must be considered was likely due to: (1) differences in user window/level preferences and related impact on visualization of target boundaries; (2) discrepancies in visualization at the superior/inferior aspect of the target. The first factor leads to systematic over- or under-estimation of the target volume. The second factor affects only SE5mm and SE3mm and leads to over- or under-estimation at superior/inferior aspect of the target. Both factors have ramification in clinical SRS treatment planning.Significant improvement in target definition was observed as the MR image resolution improved. Results imply that the highest resolution 3D MR sequences should be used to minimize potential errors in target definition, and multi-slice 2D sequence should be avoided.

A Transfer Learning Approach for Improving OAR Segmentation in the Adaptive Therapy or Retreatment of Head and Neck Cancer

<h3>Purpose/Objective(s)</h3> Deep learning-based segmentation has achieved great progress compared with traditional method. However, the generalizability is still an unsolved problem, leading to suboptimal contours for some cases that requires manual editing. A big source of error lies in the abnormal patient anatomy that was not well covered by the training data. In adaptive or retreatment settings, a well-contoured CT is available. However, the current DLSEG algorithm does not allow such information to be utilized. Our proposed solution is to use transfer learning method to tune a trained DLSEG model toward individual patient using the prior segmented scans. <h3>Materials/Methods</h3> A HN DLSEG model was trained on 2015 MICCAI Head and Neck dataset. The model was applied on HNSCC-3DCT-RT dataset from TCIA which contains 3 CT scans at pre-, mid-, and end- treatment course. The DLSEG result was compared with the ground truth contour in the study. Dice similarity coefficient (DSC), 95% Hausdorff distance (HD95), and mean surface distance (MSD) were computed for BrainStem, Mandible, Parotid, SpinalCord and OpticNerve. The standard deviations of these metrics were computed, both within the same patient (IntraPatient variation) and across patients (InterPatient variation). Then, the DLSEG was re-tuned using transfer learning on the pre- treatment CT scan of each individual patient. The re-tuned model was applied to the mid- and end- treatment CT. The segmentation accuracy was evaluated. <h3>Results</h3> All structures showed a much larger InterPatient variation than IntraPatient variation, indicating that patient specific anatomical feature is one of the main contributors to segmentation accuracy. When transfer learning is applied. Nearly all patients saw improvement in contouring accuracy. Statistically significant improvement was observed for BrainStem, Mandible, Parotid and Parotid_L. <h3>Conclusion</h3> Transfer learning is a promising approach to incorporate the existing segmentation on prior CT to improve the segmentation accuracy in the adaptive or re-treatment settings.

A Deep Learning-Based Automatic Contour Quality Assurance Pipeline for Complex Anatomy on MRI

Despite the recent significant improvement on deep learning (DL)-based auto-segmentation, review and editing on auto-segmented contours are often needed to avoid errors. Such a manual time-consuming process is not desirable, particularly for recently introduced MR-guided adaptive radiation therapy (MRgART). To eliminate this manual process, we have previously proposed automatic contour quality assurance (ACQA) methods to automatically verify contour accuracy and have demonstrated that the ACQA methods worked well for most of organs, except for complex anatomy (e.g., stomach, bowels). This study aims to develop a DL-based ACQA pipeline specifically for complex organs that have inherent anatomy variations and intensity heterogeneity on MRIs.The proposed pipeline includes the following major parts: 1) image pre-processing consisting of bias correction, normalization, intensity clipping, gas/high intensity contents filling and image cropping, 2)generation of three-channel RGB images by combining 30% core, inner shell and 4mm outer shell, 3) data augmentation through randomized 3D and 2D rotation, scaling and translation, 4) abnormal detection using statistical geometrical features extracted from ground truth contours (geo-check), and 5) training of a convolutional neural network (CNN) image classification model using samples passing geo-check. The DL auto-segmented contours of stomach on T2-weighted abdominal MRIs from 54 patients were utilized (40 for training and five-fold cross validation and 14 for independent test). A slice was labeled as inaccurate if the contour on the slice has mean distance to agreement (MDA) ≥2 mm or 95% Hausdorff distance (HD) ≥10 mm as compared to the ground truth. The labels were then manually checked to avoid mislabeling. The model performance was evaluated using area under curve (AUC) from receiver operating characteristic curve (ROC), predicting accuracy and F1 score.A total of 139,230 sample images were generated for model training and 4826 for testing. The percentage of accurate slices identified were 64% and 72%, respectively, for the training and testing datasets. The average AUC/accuracy/F1 were 0.94/0.88/0.94 for cross validation, and 0.9/0.85/0.9 and 0.92/0.88/0.92 for the DL without or with geo-check over all testing cases. For each testing patient, the accuracy/F1 score were in the range of 0.78-0.98/ 0.71-0.99 with a median value of 0.89/0.91. The execution time for a testing contour was less than 1 second on a I7 CPU and a GTX 1060 GPU.The proposed DL-ACQA method can quickly and automatically identify accurate and inaccurate contour slices after auto-segmentation of complex anatomy, e.g., stomach. With further development, this method may be integrated into current workflow, facilitating efficient and accurate segmentation in routine treatment planning or MRgART.

Real-Time Motion Tracking Based on Orthogonal Cine MRI During MR-Guided Radiation Therapy for Prostate Cancer

In contrast to radiographic imaging with implanted radio-opaque fiducials, the high soft tissue contrast and noninvasive nature of MRI are advantageous for real-time motion tracking of prostate motion. The purpose of this study is to investigate the feasibility of tracking the intrafraction prostate motion using an intensity-based template registration technique between real-time 2D cine and 3D MRI acquired during MR-guided Adaptative Radiation Therapy (MRgART).The motion tracking algorithm includes two steps: (1) creating a template by rigidly registering a set of cine images with a 3D reference MRI, and (2) tracking the target motion by registering incoming (i.e., real-time) cine with the template with a known relationship to the 3D volume. The daily 3D T2 MRI and 2D T2/T1 bFFE real-time cine MRI acquired on a 1.5T MR-Linac during MRgART for 6 prostate cancer patients were selected to test the algorithm limits. The cine images were acquired with a temporal resolution of 0.2 s for a total duration of 9-11 minutes interleaved between coronal and sagittal orientations. A 2D template image set was generated in the coronal and sagittal planes using the cine data acquired during the first 20 s registered to the 3D volume using in-plane rigid-body image registration based on a region of interest defined by the prostate contour on 3D T2 MRI. Motion tracking of the prostate was performed by registering the rest of the 2D cine series with the corresponding template. Manually delineated contours on sampled cine frames were used as the ground truth. Tracking accuracy was defined as the standard deviation of the error (SDE) between the center of mass (CM) of the ground-truth and tracked motions in 1 min. Prostate drift was evaluated by tracking the CM shift with the algorithm every 8-9 s for 9-11 min.The prostate tracking algorithm was qualitatively validated by visual assessment of 6 prostate cases. Overall, the tracking was found robust to implant-induced image artefacts in two cases, large change in bladder size in one case, and body motion in one other case. The tracking algorithm was also validated quantitatively on 2 cases for which ground truth contours were available and reliable. Mean SDE for the two cases was 0.45, 0.32, and 0.31 mm in superior-inferior, left-right, and anterior-posterior directions, respectively. A continuous motion drift of the prostate was observed over a 600 s period for one of these two cases and was accurately tracked by the technique.We have demonstrated the feasibility to track prostate motion based on real-time orthogonal cine MRI without implanting fiducials. With further investigation on more patients and following careful workflow, the real-time motion tracking technique may be implemented in MRgART to manage intrafraction motion.H. Jassar: None. A. Tai: None. X. Chen: None. T. Keiper: None. E.S. Paulson: Research Grant; Siemens Healthineers, Elekta Instruments AB.F. Lathuiliere: None. S. Bériault: None. D. Cooper: None. F. Hebert: None. L. Savard: None. S. Cloake: None. W.A. Hall: Research Grant; Elekta, National Cancer Institute, American Cancer Society. C.A. Lawton: None. A. Li: Research Grant; Elekta AB, Accuray Inc, Siemens Healthineers, Manteia Med. Honoraria; Elekta AB, Accuray Inc. Patent/License Fees/Copyright; Manteia Med.

Deep Learning Auto-Segmentation on Multi-Sequence MRI for MR-Guided Adaptive Radiation Therapy

<h3>Purpose/Objective(s)</h3> Robust auto-segmentation based on daily MRI is highly desirable for MR-guided adaptive radiation therapy (MRgART). However, it is generally challenging particularly if the daily MRI is acquired with multiple sequences or if it is for imaging complex anatomy, e.g., abdomen. To overcome these issues, we developed a deep learning algorithm to segment abdominal organs quickly and automatically on multi-sequence MRI. <h3>Materials/Methods</h3> The deep learning auto-segmentation model was developed using a modified U-net, ResUnet3D, with end-to-end mapping, and was trained with 85 daily MRI sets acquired during MRgART on a 1.5T MR-Linac for 37 patients with abdominal tumors. The daily MRI sets were motion averaged images derived from 4D MRI acquired with multiple sequences, including Fast Field Echo (FFE), Balanced Turbo Field Echo (BTFE) and TFE. Each daily MRI was paired with a CT. All MRIs were pre-processed using bias field correction and intensity standardization. The ground truth contours were created manually based on existing guidelines. For the training datasets, field-of-view extension via online random 3D elastic transform/deformation was utilized to augment and improve the quality of datasets. During the model training, a patient-based z-score normalization method was employed to standardize the multi-contrast images. Two models were trained, one for 11 organs: aorta, duodenum, kidneys, liver, pancreas, spinal cord, spleen, stomach, large bowel, and small bowel, and another for 10 organs with combined large and small bowel. Performance of the obtained models was evaluated using 11 test datasets in terms of dice similarity coefficient (DSC), mean distance to agreement (MDA) and time efficiency using an Intel Xeon CPU (2.4 GHz, 64GB RAM) hardware. <h3>Results</h3> For the challenging bowels, the auto-segmentation accuracy for the model trained with the combined bowel was better, with MDA/DSC of 2.1/0.86 versus 6.8/0.76 (large bowel) and 6.1/0.68 (small bowel) for the separate-bowel model. Average DSC for the obtained contours of aorta, left kidney, right kidney, liver, spinal cord, spleen, and stomach was 0.93, 0.93, 0.94, 0.94, 0.83, 0.92 and 0.88, respectively from both models. While no significant difference in model performance was observed with BTFE versus TFE testing datasets, a slightly improved segmentation was seen on the BTFE images for pancreas and duodenum, DSC 0.71 and 0.72. This improvement may be attributed to the enhanced contrast seen in BTFE images when compared to the TFE and FFE images. The time required to generate all organs on an MRI was 20-24 seconds. <h3>Conclusion</h3> We have developed deep learning auto-segmentation models based on multi-sequence MRI sets for abdominal organs. The models performed reasonably well. With further development using large datasets, the models may be integrated in MRgART for fast, fully automated, and accurate contouring of abdominal organs on daily MRI.

Comparing Deep Learning Based Prostate Delineation With Manual Contouring From Multiple Experts

Accurate contouring of the target is very important in radiation therapy treatment planning. Unlike most other cancers where the radiation target occupies a small area of an organ, prostate cancer treatment involves the irradiation of the entire prostate. This study evaluates the quality of prostate contours generated from a deep learning segmentation (DLSEG) model.A total of 23 prostate cancer patients who did not underwent prostatectomy was randomly selected. The original contours used clinically was carefully quality checked and revised when necessary by one expert radiation oncologist (RO) and regarded as the ground truth (GT). Two additional ROs were invited to perform manual contouring of the prostate from scratch. Auto contours (labelled as AI) generated by a research version of DLSEG was then compared with those from manual delineations. A paired t-test was employed for statistical comparison. A comparative geometric analysis of the manual and automatic delineations was performed to quantify the multi-observer delineation variability, including dice similarity coefficient (DSC), 95% Hausdorff distance (HD95), and mean surface distance (MSD).Using GT contour as reference, the contours from three additional RO achieved an average of 0.81 ± 0.04 for DSC, 0.60 ± 0.16cm and 0.21 ± 0.05cm for HD95 and MSD, respectively. In contrast, AI contours achieved an average of 0.83 ± 0.05 for DSC, 0.60 ± 0.18cm and 0.21 ± 0.07 cm for HD95 and MSD, respectively. No statistical difference was found for the delineation accuracy between the three observers (P > 0.05). AI contours were considered to have similar performance to the three human experts with all geometric metrics (P > 0.05).The DLSEG model for prostate contouring tested demonstrates similar performance in terms of accuracy compared with expert ROs.

Discrete light sheet microscopic segmentation of left ventricle using morphological tuning and active contours.

Left ventricular segmentation using cardiovascular MR scan is required for the diagnosis and further cure of cardiac diseases. Automatic systems for left ventricle segmentation are being studied for attaining more accurate results in a shorter period of time. A novel algorithm introducing discrete segmentation of left ventricle achieves an independent processing of images swiftly. The workflow consists of four segments; first, automated localization is performed on the MR image. Second, performing preprocessing intimately improves and enhances the quality of image using mean contrast adjustment. Central segmentation of endocardium and epicardium layers includes novel MTAC (Morphological tuning using active contours) segmentation algorithm that provides a perfect combination of active contours and morphological tuning to bring an adequate and desirable segmentation. The prospective snake model is a restrained progression, which takes iterations for an impulse throughout the left ventricle contours. At the end, contrast based refining overcomes minor edge problems for both outer and inner boundaries. Proposed algorithm is evaluated via Sunnybrook cardiac MR images by producing an overall average perpendicular distance 2.45 mm, an average dice matrix (endo: 91.3%; epi: 92.16%) and 91.7% dice matrix of overall endocardium and epicardium contours from ground truth contours.

RADI-12. Deep learning for automatic detection and contouring of metastatic brain tumors in stereotactic radiosurgery: a retrospective analysis with an FDA-cleared software algorithm

IntroductionArtificial intelligence-based tools can significantly impact detection and segmentation of brain metastases for stereotactic radiosurgery (SRS). VBrain is a deep learning algorithm, recently FDA-cleared, to assist in brain tumor contouring. In this study, we aimed to further validate this tool in patients treated with SRS for brain metastases at Stanford Cancer Center.MethodsWe included randomly selected patients with brain metastases treated with SRS from 2008 to 2020. Computed tomography (CT) and axial T1-weighted post-contrast magnetic resonance (MR) image data were extracted for each patient and uploaded to VBrain. Subsequent analyses compared the output contours from VBrain with the physician-defined contours used for SRS. A brain metastasis was considered “detected” when the VBrain “predicted” contours overlapped with the corresponding physician contours (“ground-truth” contours). We evaluated performance against ground-truth contours using the following metrics: lesion-wise Dice similarity coefficient (DSC), lesion-wise average Hausdorff distance (AVD), false positive count (FP), and lesion-wise sensitivity (%).ResultsWe analyzed 60 patients with 321 intact brain metastases treated over 70 SRS courses. Resection cavities were excluded from the analysis. The median (range) tumor size was 132 mm3 (7 to 24,765). Input CT scan slice thickness was 1.250 mm, and median (range) pixel resolution was 0.547 mm (0.457 to 0.977). Input MR scan median (range) slice thickness was 1.000 mm (0.940 to 2.000), and median (range) pixel resolution was 0.469 mm (0.469 to 1.094). In assessing VBrain performance, we found mean lesion-wise DSC to be 0.70, mean lesion-wise AVD to be 9.40% of lesion size (0.805 mm), mean FP to be 0.657 tumors per case, and lesion-wise sensitivity to be 84.5%.ConclusionRetrospective analysis of our brain metastases cohort using a deep learning algorithm yielded promising results. Integration of VBrain into the clinical workflow can provide further clinical and research insights.

Validation of accuracy deformable image registration contour propagation using a benchmark virtual HN phantom dataset.

Virtual anatomic phantoms offer precise voxel mapping of the variation of anatomy with ground truth deformation vector fields (DVFs). Dice similarity coefficient (DSC) and mean distance to agreement (MDA) are the standard metrics for evaluating geometric contour congruence when testing deformable registration (DIR) algorithms. A HN virtual patient phantom data set was used for a kVCT‐kVCT automatic propagation contour validation study employing the Accuray DIR algorithm. Furthermore, since TomoTherapy uses MVCT images of the relevant anatomy for adaptive monitoring, the kVCT image data set quality was transformed to an MVCT image data set quality to study intermodal kVCT‐MVCT DIR accuracy. The results of the study indicate that the Accuray DIR algorithm can be expected to autopropagate HN contours adequately, on average, within tolerances recommended by TG‐132 (DSC 0.8‐0.9, MDA within voxel width). However, contours critical to dosimetric planning should always be visually proofed for accuracy. Using standard reconstruction MVCT image quality causes slightly less, but acceptable, agreement with ground truth contours.

Automatic delineation of cardiac substructures using a region-based fully convolutional network.

Radiation dose to specific cardiac substructures, such as the atria and ventricles, has been linked to post-treatment toxicity and has shown to be more predictive of these toxicities than dose to the whole heart. A deep learning-based algorithm for automatic generation of these contours is proposed to aid in either retrospective or prospective dosimetric studies to better understand the relationship between radiation dose and toxicities. The proposed method uses a mask-scoring regional convolutional neural network (RCNN) which consists of five major subnetworks: backbone, regional proposal network (RPN), RCNN head, mask head, and mask-scoring head. Multiscale feature maps are learned from computed tomography (CT) via the backbone network. The RPN utilizes these feature maps to detect the location and region-of-interest (ROI) of all substructures, and the final three subnetworks work in series to extract structural information from these ROIs. The network is trained using 55 patient CT datasets, with 22 patients having contrast scans. Threefold cross validation (CV) is used for evaluation on 45 datasets, and a separate cohort of 10 patients are used for holdout evaluation. The proposed method is compared to a 3D UNet. The proposed method produces contours that are qualitatively similar to the ground truth contours. Quantitatively, the proposed method achieved average Dice score coefficients (DSCs) for the whole heart, chambers, great vessels, coronary arteries, the valves of the heart of 0.96, 0.94, 0.93, 0.66, and 0.77 respectively, outperforming the 3D UNet, which achieved DSCs of 0.92, 0.87, 0.88, 0.48, and 0.59 for the corresponding substructure groups. Mean surface distances (MSDs) between substructures segmented by the proposed method and the ground truth were <2mm except for the left anterior descending coronary artery and the mitral and tricuspid valves, and <5mm for all substructures. When dividing results into noncontrast and contrast datasets, the model performed statistically significantly better in terms of DSC, MSD, centroid mean distance (CMD), and volume difference for the chambers and whole heart with contrast. Notably, the presence of contrast did not statistically significantly affect coronary artery segmentation DSC or MSD. After network training, all substructures and the whole heart can be segmented on new datasets in less than 5 s. A deep learning network was trained for automatic delineation of cardiac substructures based on CT alone. The proposed method can be used as a tool to investigate the relationship between cardiac substructure dose and treatment toxicities.

Machine Learning Driven Contouring of High-Frequency Four-Dimensional Cardiac Ultrasound Data.

Automatic boundary detection of 4D ultrasound (4DUS) cardiac data is a promising yet challenging application at the intersection of machine learning and medicine. Using recently developed murine 4DUS cardiac imaging data, we demonstrate here a set of three machine learning models that predict left ventricular wall kinematics along both the endo- and epi-cardial boundaries. Each model is fundamentally built on three key features: (1) the projection of raw US data to a lower dimensional subspace, (2) a smoothing spline basis across time, and (3) a strategic parameterization of the left ventricular boundaries. Model 1 is constructed such that boundary predictions are based on individual short-axis images, regardless of their relative position in the ventricle. Model 2 simultaneously incorporates parallel short-axis image data into their predictions. Model 3 builds on the multi-slice approach of model 2, but assists predictions with a single ground-truth position at end-diastole. To assess the performance of each model, Monte Carlo cross validation was used to assess the performance of each model on unseen data. For predicting the radial distance of the endocardium, models 1, 2, and 3 yielded average R2 values of 0.41, 0.49, and 0.71, respectively. Monte Carlo simulations of the endocardial wall showed significantly closer predictions when using model 2 versus model 1 at a rate of 48.67%, and using model 3 versus model 2 at a rate of 83.50%. These finding suggest that a machine learning approach where multi-slice data are simultaneously used as input and predictions are aided by a single user input yields the most robust performance. Subsequently, we explore the how metrics of cardiac kinematics compare between ground-truth contours and predicted boundaries. We observed negligible deviations from ground-truth when using predicted boundaries alone, except in the case of early diastolic strain rate, providing confidence for the use of such machine learning models for rapid and reliable assessments of murine cardiac function. To our knowledge, this is the first application of machine learning to murine left ventricular 4DUS data. Future work will be needed to strengthen both model performance and applicability to different cardiac disease models.

Liver segmentation in abdominal CT images via auto-context neural network and self-supervised contour attention

Objective: Accurate image segmentation of the liver is a challenging problem owing to its large shape variability and unclear boundaries. Although the applications of fully convolutional neural networks (CNNs) have shown groundbreaking results, limited studies have focused on the performance of generalization. In this study, we introduce a CNN for liver segmentation on abdominal computed tomography (CT) images that focus on the performance of generalization and accuracy. Methods: To improve the generalization performance, we initially propose an auto-context algorithm in a single CNN. The proposed auto-context neural network exploits an effective high-level residual estimation to obtain the shape prior. Identical dual paths are effectively trained to represent mutual complementary features for an accurate posterior analysis of a liver. Further, we extend our network by employing a self-supervised contour scheme. We trained sparse contour features by penalizing the ground-truth contour to focus more contour attentions on the failures. Results: We used 180 abdominal CT images for training and validation. Two-fold cross-validation is presented for a comparison with the state-of-the-art neural networks. The experimental results show that the proposed network results in better accuracy when compared to the state-of-the-art networks by reducing 10.31% of the Hausdorff distance. Novel multiple N-fold cross-validations are conducted to show the best performance of generalization of the proposed network. Conclusion and significance: The proposed method minimized the error between training and test images more than any other modern neural networks. Moreover, the contour scheme was successfully employed in the network by introducing a self-supervising metric.