ABSTRACT
BACKGROUND AND OBJECTIVES: Bio-medical image segmentation models typically attempt to predict one segmentation that resembles a ground-truth structure as closely as possible. However, as medical images are not perfect representations of anatomy, obtaining this ground truth is not possible. A surrogate commonly used is to have multiple expert observers define the same structure for a dataset. When multiple observers define the same structure on the same image there can be significant differences depending on the structure, image quality/modality and the region being defined. It is often desirable to estimate this type of aleatoric uncertainty in a segmentation model to help understand the region in which the true structure is likely to be positioned. Furthermore, obtaining these datasets is resource intensive so training such models using limited data may be required. With a small dataset size, differing patient anatomy is likely not well represented causing epistemic uncertainty which should also be estimated so it can be determined for which cases the model is effective or not. METHODS: We use a 3D probabilistic U-Net to train a model from which several segmentations can be sampled to estimate the range of uncertainty seen between multiple observers. To ensure that regions where observers disagree most are emphasised in model training, we expand the Generalised Evidence Lower Bound (ELBO) with a Constrained Optimisation (GECO) loss function with an additional contour loss term to give attention to this region. Ensemble and Monte-Carlo dropout (MCDO) uncertainty quantification methods are used during inference to estimate model confidence on an unseen case. We apply our methodology to two radiotherapy clinical trial datasets, a gastric cancer trial (TOPGEAR, TROG 08.08) and a post-prostatectomy prostate cancer trial (RAVES, TROG 08.03). Each dataset contains only 10 cases each for model development to segment the clinical target volume (CTV) which was defined by multiple observers on each case. An additional 50 cases are available as a hold-out dataset for each trial which had only one observer define the CTV structure on each case. Up to 50 samples were generated using the probabilistic model for each case in the hold-out dataset. To assess performance, each manually defined structure was matched to the closest matching sampled segmentation based on commonly used metrics. RESULTS: The TOPGEAR CTV model achieved a Dice Similarity Coefficient (DSC) and Surface DSC (sDSC) of 0.7 and 0.43 respectively with the RAVES model achieving 0.75 and 0.71 respectively. Segmentation quality across cases in the hold-out datasets was variable however both the ensemble and MCDO uncertainty estimation approaches were able to accurately estimate model confidence with a p-value < 0.001 for both TOPGEAR and RAVES when comparing the DSC using the Pearson correlation coefficient. CONCLUSIONS: We demonstrated that training auto-segmentation models which can estimate aleatoric and epistemic uncertainty using limited datasets is possible. Having the model estimate prediction confidence is important to understand for which unseen cases a model is likely to be useful.
ABSTRACT
Cervical cancer is a common cancer in women globally, with treatment usually involving radiation therapy (RT). Accurate segmentation for the tumour site and organ-at-risks (OARs) could assist in the reduction of treatment side effects and improve treatment planning efficiency. Cervical cancer Magnetic Resonance Imaging (MRI) segmentation is challenging due to a limited amount of training data available and large inter- and intra- patient shape variation for OARs. The proposed Masked-Net consists of a masked encoder within the 3D U-Net to account for the large shape variation within the dataset, with additional dilated layers added to improve segmentation performance. A new loss function was introduced to consider the bounding box loss during training with the proposed Masked-Net. Transfer learning from a male pelvis MRI data with a similar field of view was included. The approaches were compared to the 3D U-Net which was widely used in MRI image segmentation. The data used consisted of 52 volumes obtained from 23 patients with stage IB to IVB cervical cancer across a maximum of 7 weeks of RT with manually contoured labels including the bladder, cervix, gross tumour volume, uterus and rectum. The model was trained and tested with a 5-fold cross validation. Outcomes were evaluated based on the Dice Similarity Coefficients (DSC), the Hausdorff Distance (HD) and the Mean Surface Distance (MSD). The proposed method accounted for the small dataset, large variations in OAR shape and tumour sizes with an average DSC, HD and MSD for all anatomical structures of 0.790, 30.19mm and 3.15mm respectively.
ABSTRACT
Objectives.Contouring similarity metrics are often used in studies of inter-observer variation and automatic segmentation but do not provide an assessment of clinical impact. This study focused on post-prostatectomy radiotherapy and aimed to (1) identify if there is a relationship between variations in commonly used contouring similarity metrics and resulting dosimetry and (2) identify the variation in clinical target volume (CTV) contouring that significantly impacts dosimetry.Approach.The study retrospectively analysed CT scans of 10 patients from the TROG 08.03 RAVES trial. The CTV, rectum, and bladder were contoured independently by three experienced observers. Using these contours reference simultaneous truth and performance level estimation (STAPLE) volumes were established. Additional CTVs were generated using an atlas algorithm based on a single benchmark case with 42 manual contours. Volumetric-modulated arc therapy (VMAT) treatment plans were generated for the observer, atlas, and reference volumes. The dosimetry was evaluated using radiobiological metrics. Correlations between contouring similarity and dosimetry metrics were calculated using Spearman coefficient (Γ). To access impact of variations in planning target volume (PTV) margin, the STAPLE PTV was uniformly contracted and expanded, with plans created for each PTV volume. STAPLE dose-volume histograms (DVHs) were exported for plans generated based on the contracted/expanded volumes, and dose-volume metrics assessed.Mainresults. The study found no strong correlations between the considered similarity metrics and modelled outcomes. Moderate correlations (0.5 <Γ< 0.7) were observed for Dice similarity coefficient, Jaccard, and mean distance to agreement metrics and rectum toxicities. The observations of this study indicate a tendency for variations in CTV contraction/expansion below 5 mm to result in minor dosimetric impacts.Significance. Contouring similarity metrics must be used with caution when interpreting them as indicators of treatment plan variation. For post-prostatectomy VMAT patients, this work showed variations in contours with an expansion/contraction of less than 5 mm did not lead to notable dosimetric differences, this should be explored in a larger dataset to assess generalisability.
Subject(s)
Prostatic Neoplasms , Radiotherapy, Intensity-Modulated , Male , Humans , Prostate , Prostatic Neoplasms/diagnostic imaging , Prostatic Neoplasms/radiotherapy , Prostatic Neoplasms/surgery , Radiotherapy Planning, Computer-Assisted/methods , Retrospective Studies , Radiotherapy, Intensity-Modulated/methods , Radiotherapy Dosage , Treatment OutcomeABSTRACT
Background and Purpose: Addressing the need for accurate dose calculation in MRI-only radiotherapy, the generation of synthetic Computed Tomography (sCT) from MRI has emerged. Deep learning (DL) techniques, have shown promising results in achieving high sCT accuracies. However, existing sCT synthesis methods are often center-specific, posing a challenge to their generalizability. To overcome this limitation, recent studies have proposed approaches, such as multicenter training . Material and methods: The purpose of this work was to propose a multicenter sCT synthesis by DL, using a 2D cycle-GAN on 128 prostate cancer patients, from four different centers. Four cases were compared: monocenter cases, monocenter training and test on another center, multicenter trainings and a test on a center not included in the training and multicenter trainings with an included center in the test. Trainings were performed using 20 patients. sCT accuracy evaluation was performed using Mean Absolute Error, Mean Error and Peak-Signal-to-Noise-Ratio. Dose accuracy was assessed with gamma index and Dose Volume Histogram comparison. Results: Qualitative, quantitative and dose results show that the accuracy of sCTs for monocenter trainings and multicenter trainings using a seen center in the test did not differ significantly. However, when the test involved an unseen center, the sCT quality was inferior. Conclusions: The aim of this work was to propose generalizable multicenter training for MR-to-CT synthesis. It was shown that only a few data from one center included in the training cohort allows sCT accuracy equivalent to a monocenter study.
ABSTRACT
Introduction: For radiotherapy based solely on magnetic resonance imaging (MRI), generating synthetic computed tomography scans (sCT) from MRI is essential for dose calculation. The use of deep learning (DL) methods to generate sCT from MRI has shown encouraging results if the MRI images used for training the deep learning network and the MRI images for sCT generation come from the same MRI device. The objective of this study was to create and evaluate a generic DL model capable of generating sCTs from various MRI devices for prostate radiotherapy. Materials and methods: In total, 90 patients from three centers (30 CT-MR prostate pairs/center) underwent treatment using volumetric modulated arc therapy for prostate cancer (PCa) (60 Gy in 20 fractions). T2 MRI images were acquired in addition to computed tomography (CT) images for treatment planning. The DL model was a 2D supervised conditional generative adversarial network (Pix2Pix). Patient images underwent preprocessing steps, including nonrigid registration. Seven different supervised models were trained, incorporating patients from one, two, or three centers. Each model was trained on 24 CT-MR prostate pairs. A generic model was trained using patients from all three centers. To compare sCT and CT, the mean absolute error in Hounsfield units was calculated for the entire pelvis, prostate, bladder, rectum, and bones. For dose analysis, mean dose differences of D 99% for CTV, V 95% for PTV, Dmax for rectum and bladder, and 3D gamma analysis (local, 1%/1 mm) were calculated from CT and sCT. Furthermore, Wilcoxon tests were performed to compare the image and dose results obtained with the generic model to those with the other trained models. Results: Considering the image results for the entire pelvis, when the data used for the test comes from the same center as the data used for training, the results were not significantly different from the generic model. Absolute dose differences were less than 1 Gy for the CTV D 99% for every trained model and center. The gamma analysis results showed nonsignificant differences between the generic and monocentric models. Conclusion: The accuracy of sCT, in terms of image and dose, is equivalent to whether MRI images are generated using the generic model or the monocentric model. The generic model, using only eight MRI-CT pairs per center, offers robust sCT generation, facilitating PCa MRI-only radiotherapy for routine clinical use.
ABSTRACT
BACKGROUND: Point-of-care lung ultrasound (LUS) allows real-time patient scanning to help diagnose pleural effusion (PE) and plan further investigation and treatment. LUS typically requires training and experience from the clinician to accurately interpret the images. To address this limitation, we previously demonstrated a deep-learning model capable of detecting the presence of PE on LUS at an accuracy greater than 90%, when compared to an experienced LUS operator. METHODS: This follow-up study aimed to develop a deep-learning model to provide segmentations for PE in LUS. Three thousand and forty-one LUS images from twenty-four patients diagnosed with PE were selected for this study. Two LUS experts provided the ground truth for training by reviewing and segmenting the images. The algorithm was then trained using ten-fold cross-validation. Once training was completed, the algorithm segmented a separate subset of patients. RESULTS: Comparing the segmentations, we demonstrated an average Dice Similarity Coefficient (DSC) of 0.70 between the algorithm and experts. In contrast, an average DSC of 0.61 was observed between the experts. CONCLUSION: In summary, we showed that the trained algorithm achieved a comparable average DSC at PE segmentation. This represents a promising step toward developing a computational tool for accurately augmenting PE diagnosis and treatment.
Subject(s)
Deep Learning , Pleural Effusion , Humans , Follow-Up Studies , Algorithms , Lung/diagnostic imaging , Pleural Effusion/diagnostic imagingABSTRACT
Automatically generating a report from a patient's Chest X-rays (CXRs) is a promising solution to reducing clinical workload and improving patient care. However, current CXR report generators-which are predominantly encoder-to-decoder models-lack the diagnostic accuracy to be deployed in a clinical setting. To improve CXR report generation, we investigate warm starting the encoder and decoder with recent open-source computer vision and natural language processing checkpoints, such as the Vision Transformer (ViT) and PubMedBERT. To this end, each checkpoint is evaluated on the MIMIC-CXR and IU X-ray datasets. Our experimental investigation demonstrates that the Convolutional vision Transformer (CvT) ImageNet-21K and the Distilled Generative Pre-trained Transformer 2 (DistilGPT2) checkpoints are best for warm starting the encoder and decoder, respectively. Compared to the state-of-the-art (M2 Transformer Progressive), CvT2DistilGPT2 attained an improvement of 8.3% for CE F-1, 1.8% for BLEU-4, 1.6% for ROUGE-L, and 1.0% for METEOR. The reports generated by CvT2DistilGPT2 have a higher similarity to radiologist reports than previous approaches. This indicates that leveraging warm starting improves CXR report generation. Code and checkpoints for CvT2DistilGPT2 are available at https://github.com/aehrc/cvt2distilgpt2.
Subject(s)
Natural Language Processing , Workload , Humans , X-RaysABSTRACT
Radiation therapy is moving from CT based to MRI guided planning, particularly for soft tissue anatomy. An important requirement of this new workflow is the generation of synthetic-CT (sCT) from MRI to enable treatment dose calculations. Automatic methods to determine the acceptable range of CT Hounsfield Unit (HU) uncertainties to avoid dose distribution errors is thus a key step toward safe MRI-only radiotherapy. This work has analysed the effects of controlled errors introduced in CT scans on the delivered radiation dose for prostate cancer patients. Spearman correlation coefficient has been computed, and a global sensitivity analysis performed following the Morris screening method. This allows the classification of different error factors according to their impact on the dose at the isocentre. sCT HU estimation errors in the bladder appeared to be the least influential factor, and sCT quality assessment should not only focus on organs surrounding the radiation target, as errors in other soft tissue may significantly impact the dose in the target volume. This methodology links dose and intensity-based metrics, and is the first step to define a threshold of acceptability of HU uncertainties for accurate dose planning.
Subject(s)
Prostate , Prostatic Neoplasms , Male , Humans , Prostate/diagnostic imaging , Tomography, X-Ray Computed/methods , Prostatic Neoplasms/diagnostic imaging , Prostatic Neoplasms/radiotherapy , Urinary Bladder , Magnetic Resonance Imaging/methodsABSTRACT
Combined magnetic resonance imaging (MRI) and positron emission tomography/computed tomography (PET/CT) may enhance diagnosis, aid surgical planning and intra-operative orientation for prostate biopsy and radical prostatectomy. Although PET-MRI may provide these benefits, PET-MRI machines are not widely available. Image fusion of Prostate specific membrane antigen PET/CT and MRI acquired separately may be a suitable clinical alternative. This study compares CT-MR registration algorithms for urological prostate cancer care. Paired whole-pelvis MR and CT scan data were used (n = 20). A manual prostate CTV contour was performed independently on each patients MR and CT image. A semi-automated rigid-, automated rigid- and automated non-rigid registration technique was applied to align the MR and CT data. Dice Similarity Index (DSI), 95% Hausdorff distance (95%HD) and average surface distance (ASD) measures were used to assess the closeness of the manual and registered contours. The automated non-rigid approach had a significantly improved performance compared to the automated rigid- and semi-automated rigid-registration, having better average scores and decreased spread for the DSI, 95%HD and ASD (all p < 0.001). Additionally, the automated rigid approach had similar significantly improved performance compared to the semi-automated rigid registration across all accuracy metrics observed (all p < 0.001). Overall, all registration techniques studied here demonstrated sufficient accuracy for exploring their clinical use. While the fully automated non-rigid registration algorithm in the present study provided the most accurate registration, the semi-automated rigid registration is a quick, feasible, and accessible method to perform image registration for prostate cancer care by urologists and radiation oncologists now.
Subject(s)
Prostate , Prostatic Neoplasms , Male , Humans , Prostate/diagnostic imaging , Prostate/surgery , Positron Emission Tomography Computed Tomography , Tomography, X-Ray Computed/methods , Prostatic Neoplasms/diagnostic imaging , Prostatic Neoplasms/surgery , PelvisABSTRACT
Background and purpose: Magnetic Resonance Imaging (MRI)-only planning workflows offer many advantages but raises challenges regarding image guidance. The study aimed to assess the viability of MRI to Cone Beam Computed Tomography (CBCT) based image guidance for MRI-only planning treatment workflows. Materials and methods: An MRI matching training package was developed. Ten radiation therapists, with a range of clinical image guidance experience and experience with MRI, completed the training package prior to matching assessment. The matching assessment was performed on four match regions: prostate gold seed, prostate soft tissue, rectum/anal canal and gynaecological. Each match region consisted of five patients, with three CBCTs per patient, resulting in fifteen CBCTs for each match region. The ten radiation therapists performed the CBCT image matching to CT and to MRI for all regions and recorded the match values. Results: The median inter-observer variation for MRI-CBCT matching and CT-CBCT matching for all regions were within 2 mm and 1 degree. There was no statistically significant association in the inter-observer variation in mean match values and radiation therapist image guidance experience levels. There was no statistically significant association in inter-observer variation in mean match values for MRI experience levels for prostate soft tissue and gynaecological match regions, while there was a statistically significant difference for prostate gold seed and rectum match regions. Conclusion: The results of this study support the concept that with focussed training, an MRI to CBCT image guidance approach can be successfully implemented in a clinical planning workflow.
ABSTRACT
BACKGROUND AND PURPOSE: Previous studies on automatic delineation quality assurance (QA) have mostly focused on CT-based planning. As MRI-guided radiotherapy is increasingly utilized in prostate cancer treatment, there is a need for more research on MRI-specific automatic QA. This work proposes a clinical target volume (CTV) delineation QA framework based on deep learning (DL) for MRI-guided prostate radiotherapy. MATERIALS AND METHODS: The proposed workflow utilized a 3D dropblock ResUnet++ (DB-ResUnet++) to generate multiple segmentation predictions via Monte Carlo dropout which were used to compute an average delineation and area of uncertainty. A logistic regression (LR) classifier was employed to classify the manual delineation as pass or discrepancy based on the spatial association between the manual delineation and the network's outputs. This approach was evaluated on a multicentre MRI-only prostate radiotherapy dataset and compared with our previously published QA framework based on AN-AG Unet. RESULTS: The proposed framework achieved an area under the receiver operating curve (AUROC) of 0.92, a true positive rate (TPR) of 0.92 and a false positive rate of 0.09 with an average processing time per delineation of 1.3 min. Compared with our previous work using AN-AG Unet, this method generated fewer false positive detections at the same TPR with a much faster processing speed. CONCLUSION: To the best of our knowledge, this is the first study to propose an automatic delineation QA tool using DL with uncertainty estimation for MRI-guided prostate radiotherapy, which can potentially be used for reviewing prostate CTV delineation in multicentre clinical trials.
Subject(s)
Deep Learning , Prostatic Neoplasms , Radiotherapy, Image-Guided , Humans , Male , Quality Assurance, Health Care , Magnetic Resonance Imaging , Uncertainty , Prostatic Neoplasms/diagnostic imaging , Prostatic Neoplasms/radiotherapyABSTRACT
Radiotherapy treatment planning based only on magnetic resonance imaging (MRI) has become clinically achievable. Though computed tomography (CT) is the gold standard for radiotherapy imaging, directly providing the electron density values needed for planning calculations, MRI has superior soft tissue visualisation to guide treatment planning decisions and optimisation. MRI-only planning removes the need for the CT scan, but requires generation of a substitute/synthetic/pseudo CT (sCT) for electron density information. Shortening the MRI imaging time would improve patient comfort and reduce the likelihood of motion artefacts. A volunteer study was previously carried out to investigate and optimise faster MRI sequences for a hybrid atlas-voxel conversion to sCT for prostate treatment planning. The aim of this follow-on study was to clinically validate the performance of the new optimised sequence for sCT generation in a treated MRI-only prostate patient cohort. 10 patients undergoing MRI-only treatment were scanned on a Siemens Skyra 3T MRI as part of the MRI-only sub-study of the NINJA clinical trial (ACTRN12618001806257). Two sequences were used, the standard 3D T2-weighted SPACE sequence used for sCT conversion which has been previously validated against CT, and a modified fast SPACE sequence, selected based on the volunteer study. Both were used to generate sCT scans. These were then compared to evaluate the fast sequence conversion for anatomical and dosimetric accuracy against the clinically approved treatment plans. The average Mean Absolute Error (MAE) for the body was 14.98 ± 2.35 HU, and for bone was 40.77 ± 5.51 HU. The external volume contour comparison produced a Dice Similarity Coefficient (DSC) of at least 0.976, and an average of 0.985 ± 0.004, and the bony anatomy contour comparison a DSC of at least 0.907, and an average of 0.950 ± 0.018. The fast SPACE sCT agreed with the gold standard sCT within an isocentre dose of -0.28% ± 0.16% and an average gamma pass rate of 99.66% ± 0.41% for a 1%/1 mm gamma tolerance. In this clinical validation study, the fast sequence, which reduced the required imaging time by approximately a factor of 4, produced an sCT with similar clinical dosimetric results compared to the standard sCT, demonstrating its potential for clinical use for treatment planning.
Subject(s)
Prostate , Radiotherapy Planning, Computer-Assisted , Humans , Male , Magnetic Resonance Imaging/methods , Pelvis , Radiotherapy Planning, Computer-Assisted/methods , Tomography, X-Ray Computed/methodsABSTRACT
Distal radius fractures (DRFs) are one of the most common types of wrist fracture and can be subdivided into intra- and extra-articular fractures. Compared with extra-articular DRFs which spare the joint surface, intra-articular DRFs extend to the articular surface and can be more difficult to treat. Identification of articular involvement can provide valuable information about the characteristics of fracture patterns. In this study, a two-stage ensemble deep learning framework was proposed to differentiate intra- and extra-articular DRFs automatically on posteroanterior (PA) view wrist X-rays. The framework firstly detects the distal radius region of interest (ROI) using an ensemble model of YOLOv5 networks, which imitates the clinicians' search pattern of zooming in on relevant regions to assess abnormalities. Secondly, an ensemble model of EfficientNet-B3 networks classifies the fractures in the detected ROIs into intra- and extra-articular. The framework achieved an area under the receiver operating characteristic curve of 0.82, an accuracy of 0.81, a true positive rate of 0.83 and a false positive rate of 0.27 (specificity of 0.73) for differentiating intra- from extra-articular DRFs. This study has demonstrated the potential in automatic DRF characterization using deep learning on clinically acquired wrist radiographs and can serve as a baseline for further research in incorporating multi-view information for fracture classification.
Subject(s)
Deep Learning , Intra-Articular Fractures , Radius Fractures , Wrist Fractures , Humans , Radius Fractures/diagnostic imaging , Intra-Articular Fractures/diagnostic imaging , RadiographyABSTRACT
Radiotherapy for thoracic and breast tumours is associated with a range of cardiotoxicities. Emerging evidence suggests cardiac substructure doses may be more predictive of specific outcomes, however, quantitative data necessary to develop clinical planning constraints is lacking. Retrospective analysis of patient data is required, which relies on accurate segmentation of cardiac substructures. In this study, a novel model was designed to deliver reliable, accurate, and anatomically consistent segmentation of 18 cardiac substructures on computed tomography (CT) scans. Thirty manually contoured CT scans were included. The proposed multi-stage method leverages deep learning (DL), multi-atlas mapping, and geometric modelling to automatically segment the whole heart, cardiac chambers, great vessels, heart valves, coronary arteries, and conduction nodes. Segmentation performance was evaluated using the Dice similarity coefficient (DSC), mean distance to agreement (MDA), Hausdorff distance (HD), and volume ratio. Performance was reliable, with no errors observed and acceptable variation in accuracy between cases, including in challenging cases with imaging artefacts and atypical patient anatomy. The median DSC range was 0.81-0.93 for whole heart and cardiac chambers, 0.43-0.76 for great vessels and conduction nodes, and 0.22-0.53 for heart valves. For all structures the median MDA was below 6 mm, median HD ranged 7.7-19.7 mm, and median volume ratio was close to one (0.95-1.49) for all structures except the left main coronary artery (2.07). The fully automatic algorithm takes between 9 and 23 min per case. The proposed fully-automatic method accurately delineates cardiac substructures on radiotherapy planning CT scans. Robust and anatomically consistent segmentations, particularly for smaller structures, represents a major advantage of the proposed segmentation approach. The open-source software will facilitate more precise evaluation of cardiac doses and risks from available clinical datasets.
Subject(s)
Heart , Image Processing, Computer-Assisted , Humans , Retrospective Studies , Image Processing, Computer-Assisted/methods , Heart/diagnostic imaging , Tomography, X-Ray Computed , AlgorithmsABSTRACT
PURPOSE: To provide a metric that reflects the dosimetric utility of the synthetic CT (sCT) and can be rapidly determined. METHODS: Retrospective CT and atlas-based sCT of 62 (53 IMRT and 9 VMAT) prostate cancer patients were used. For image similarity measurements, the sCT and reference CT (rCT) were aligned using clinical registration parameters. Conventional image similarity metrics including the mean absolute error (MAE) and mean error (ME) were calculated. The water equivalent depth (WED) was automatically determined for each patient on the rCT and sCT as the distance from the skin surface to the treatment plan isocentre at 36 equidistant gantry angles, and the mean WED difference (ΔWED¯) between the two scans was calculated. Doses were calculated on each scan pair for the clinical plan in the treatment planning system. The image similarity measurements and ΔWED¯ were then compared to the isocentre dose difference (ΔDiso) between the two scans. RESULTS: While no particular relationship to dose was observed for the other image similarity metrics, the ME results showed a linear trend against ΔDiso with R2 = 0.6, and the 95 % prediction interval for ΔDiso between -1.2 and 1 %. The ΔWED¯ results showed an improved linear trend (R2 = 0.8) with a narrower 95 % prediction interval from -0.8 % to 0.8 %. CONCLUSION: ΔWED¯ highly correlates with ΔDiso for the reference and synthetic CT scans. This is easy to calculate automatically and does not require time-consuming dose calculations. Therefore, it can facilitate the process of developing and evaluating new sCT generation algorithms.
Subject(s)
Magnetic Resonance Imaging , Radiotherapy Planning, Computer-Assisted , Male , Humans , Magnetic Resonance Imaging/methods , Radiotherapy Planning, Computer-Assisted/methods , Radiotherapy Dosage , Tomography, X-Ray Computed/methods , AlgorithmsABSTRACT
The quality assurance of synthetic CT (sCT) is crucial for safe clinical transfer to an MRI-only radiotherapy planning workflow. The aim of this work is to propose a population-based process assessing local errors in the generation of sCTs and their impact on dose distribution. For the analysis to be anatomically meaningful, a customized interpatient registration method brought the population data to the same coordinate system. Then, the voxel-based process was applied on two sCT generation methods: a bulk-density method and a generative adversarial network. The CT and MRI pairs of 39 patients treated by radiotherapy for prostate cancer were used for sCT generation, and 26 of them with delineated structures were selected for analysis. Voxel-wise errors in sCT compared to CT were assessed for image intensities and dose calculation, and a population-based statistical test was applied to identify the regions where discrepancies were significant. The cumulative histograms of the mean absolute dose error per volume of tissue were computed to give a quantitative indication of the error for each generation method. Accurate interpatient registration was achieved, with mean Dice scores higher than 0.91 for all organs. The proposed method produces three-dimensional maps that precisely show the location of the major discrepancies for both sCT generation methods, highlighting the heterogeneity of image and dose errors for sCT generation methods from MRI across the pelvic anatomy. Hence, this method provides additional information that will assist with both sCT development and quality control for MRI-based planning radiotherapy.
ABSTRACT
Our automated deep learning-based approach identifies consolidation/collapse in LUS images to aid in the identification of late stages of COVID-19 induced pneumonia, where consolidation/collapse is one of the possible associated pathologies. A common challenge in training such models is that annotating each frame of an ultrasound video requires high labelling effort. This effort in practice becomes prohibitive for large ultrasound datasets. To understand the impact of various degrees of labelling precision, we compare labelling strategies to train fully supervised models (frame-based method, higher labelling effort) and inaccurately supervised models (video-based methods, lower labelling effort), both of which yield binary predictions for LUS videos on a frame-by-frame level. We moreover introduce a novel sampled quaternary method which randomly samples only 10% of the LUS video frames and subsequently assigns (ordinal) categorical labels to all frames in the video based on the fraction of positively annotated samples. This method outperformed the inaccurately supervised video-based method and more surprisingly, the supervised frame-based approach with respect to metrics such as precision-recall area under curve (PR-AUC) and F1 score, despite being a form of inaccurate learning. We argue that our video-based method is more robust with respect to label noise and mitigates overfitting in a manner similar to label smoothing. The algorithm was trained using a ten-fold cross validation, which resulted in a PR-AUC score of 73% and an accuracy of 89%. While the efficacy of our classifier using the sampled quaternary method significantly lowers the labelling effort, it must be verified on a larger consolidation/collapse dataset, our proposed classifier using the sampled quaternary video-based method is clinically comparable with trained experts' performance.
Subject(s)
COVID-19 , Deep Learning , Humans , COVID-19/diagnostic imaging , Ultrasonography/methods , Algorithms , Lung/diagnostic imagingABSTRACT
Direct automatic segmentation of objects in 3D medical imaging, such as magnetic resonance (MR) imaging, is challenging as it often involves accurately identifying multiple individual structures with complex geometries within a large volume under investigation. Most deep learning approaches address these challenges by enhancing their learning capability through a substantial increase in trainable parameters within their models. An increased model complexity will incur high computational costs and large memory requirements unsuitable for real-time implementation on standard clinical workstations, as clinical imaging systems typically have low-end computer hardware with limited memory and CPU resources only. This paper presents a compact convolutional neural network (CAN3D) designed specifically for clinical workstations and allows the segmentation of large 3D Magnetic Resonance (MR) images in real-time. The proposed CAN3D has a shallow memory footprint to reduce the number of model parameters and computer memory required for state-of-the-art performance and maintain data integrity by directly processing large full-size 3D image input volumes with no patches required. The proposed architecture significantly reduces computational costs, especially for inference using the CPU. We also develop a novel loss function with extra shape constraints to improve segmentation accuracy for imbalanced classes in 3D MR images. Compared to state-of-the-art approaches (U-Net3D, improved U-Net3D and V-Net), CAN3D reduced the number of parameters up to two orders of magnitude and achieved much faster inference, up to 5 times when predicting with a standard commercial CPU (instead of GPU). For the open-access OAI-ZIB knee MR dataset, in comparison with manual segmentation, CAN3D achieved Dice coefficient values of (mean = 0.87 ± 0.02 and 0.85 ± 0.04) with mean surface distance errors (mean = 0.36 ± 0.32 mm and 0.29 ± 0.10 mm) for imbalanced classes such as (femoral and tibial) cartilage volumes respectively when training volume-wise under only 12G video memory. Similarly, CAN3D demonstrated high accuracy and efficiency on a pelvis 3D MR imaging dataset for prostate cancer consisting of 211 examinations with expert manual semantic labels (bladder, body, bone, rectum, prostate) now released publicly for scientific use as part of this work.
Subject(s)
Image Processing, Computer-Assisted , Imaging, Three-Dimensional , Humans , Male , Image Processing, Computer-Assisted/methods , Imaging, Three-Dimensional/methods , Neural Networks, Computer , Magnetic Resonance Imaging/methods , ProstateABSTRACT
PURPOSE: Previous work on Magnetic Resonance Imaging (MRI) only planning has been applied to limited treatment regions with a focus on male anatomy. This research aimed to validate the use of a hybrid multi-atlas synthetic computed tomography (sCT) generation technique from a MRI, using a female and male atlas, for MRI only radiation therapy treatment planning of rectum, anal canal, cervix and endometrial malignancies. PATIENTS AND METHODS: Forty patients receiving radiation treatment for a range of pelvic malignancies, were separated into male (n = 20) and female (n = 20) cohorts for the creation of gender specific atlases. A multi-atlas local weighted voting method was used to generate a sCT from a T1-weighted VIBE DIXON MRI sequence. The original treatment plans were copied from the CT scan to the corresponding sCT for dosimetric validation. RESULTS: The median percentage dose difference between the treatment plan on the CT and sCT at the ICRU reference point for the male cohort was - 0.4% (IQR of 0 to - 0.6), and - 0.3% (IQR of 0 to - 0.6) for the female cohort. The mean gamma agreement for both cohorts was > 99% for criteria of 3%/2 mm and 2%/2 mm. With dose criteria of 1%/1 mm, the pass rate was higher for the male cohort at 96.3% than the female cohort at 93.4%. MRI to sCT anatomical agreement for bone and body delineated contours was assessed, with a resulting Dice score of 0.91 ± 0.2 (mean ± 1 SD) and 0.97 ± 0.0 for the male cohort respectively; and 0.96 ± 0.0 and 0.98 ± 0.0 for the female cohort respectively. The mean absolute error in Hounsfield units (HUs) within the entire body for the male and female cohorts was 59.1 HU ± 7.2 HU and 53.3 HU ± 8.9 HU respectively. CONCLUSIONS: A multi-atlas based method for sCT generation can be applied to a standard T1-weighted MRI sequence for male and female pelvic patients. The implications of this study support MRI only planning being applied more broadly for both male and female pelvic sites. Trial registration This trial was registered in the Australian New Zealand Clinical Trials Registry (ANZCTR) ( www.anzctr.org.au ) on 04/10/2017. Trial identifier ACTRN12617001406392.
Subject(s)
Magnetic Resonance Imaging , Radiotherapy Planning, Computer-Assisted , Rectal Diseases/radiotherapy , Tomography, X-Ray Computed , Uterine Neoplasms/radiotherapy , Adult , Aged , Aged, 80 and over , Female , Humans , Male , Middle Aged , Radiotherapy DosageABSTRACT
Background and purpose: Radiation therapy (RT) is commonly indicated for treatment of prostate cancer (PC). Biologicallyoptimised RT for PC may improve disease-free survival. This requires accurate spatial localisation and characterisation of tumour lesions. We aimed to generate a statistical, voxelised biological model to complement in vivomultiparametric MRI data to facilitate biologically-optimised RT. Material and methods: Ex vivo prostate MRI and histopathological imaging were acquired for 63 PC patients. These data were co-registered to derive three-dimensional distributions of graded tumour lesions and cell density. Novel registration processes were used to map these data to a common reference geometry. Voxelised statistical models of tumour probability and cell density were generated to create the PC biological atlas. Cell density models were analysed using the Kullback-Leibler divergence to compare normal vs. lognormal approximations to empirical data. Results: A reference geometry was constructed using ex vivo MRI space, patient data were deformably registered using a novel anatomy-guided process. Substructure correspondence was maintained using peripheral zone definitions to address spatial variability in prostate anatomy between patients. Three distinct approaches to interpolation were designed to map contours, tumour annotations and cell density maps from histology into ex vivo MRI space. Analysis suggests a log-normal model provides a more consistent representation of cell density when compared to a linear-normal model. Conclusion: A biological model has been created that combines spatial distributions of tumour characteristics from a population into three-dimensional, voxelised, statistical models. This tool will be used to aid the development of biologically-optimised RT for PC patients.