Average Dice Similarity Coefficient Research Articles

Evaluation of multiple-vendor AI autocontouring solutions

BackgroundMultiple artificial intelligence (AI)-based autocontouring solutions have become available, each promising high accuracy and time savings compared with manual contouring. Before implementing AI-driven autocontouring into clinical practice, three commercially available CT-based solutions were evaluated.Materials and methodsThe following solutions were evaluated in this work: MIM-ProtégéAI+ (MIM), Radformation-AutoContour (RAD), and Siemens-DirectORGANS (SIE). Sixteen organs were identified that could be contoured by all solutions. For each organ, ten patients that had manually generated contours approved by the treating physician (AP) were identified, totaling forty-seven different patients. CT scans in the supine position were acquired using a Siemens-SOMATOMgo 64-slice helical scanner and used to generate autocontours. Physician scoring of contour accuracy was performed by at least three physicians using a five-point Likert scale. Dice similarity coefficient (DSC), Hausdorff distance (HD) and mean distance to agreement (MDA) were calculated comparing AI contours to “ground truth” AP contours.ResultsThe average physician score ranged from 1.00, indicating that all physicians reviewed the contour as clinically acceptable with no modifications necessary, to 3.70, indicating changes are required and that the time taken to modify the structures would likely take as long or longer than manually generating the contour. When averaged across all sixteen structures, the AP contours had a physician score of 2.02, MIM 2.07, RAD 1.96 and SIE 1.99. DSC ranged from 0.37 to 0.98, with 41/48 (85.4%) contours having an average DSC ≥ 0.7. Average HD ranged from 2.9 to 43.3 mm. Average MDA ranged from 0.6 to 26.1 mm.ConclusionsThe results of our comparison demonstrate that each vendor’s AI contouring solution exhibited capabilities similar to those of manual contouring. There were a small number of cases where unusual anatomy led to poor scores with one or more of the solutions. The consistency and comparable performance of all three vendors’ solutions suggest that radiation oncology centers can confidently choose any of the evaluated solutions based on individual preferences, resource availability, and compatibility with their existing clinical workflows. Although AI-based contouring may result in high-quality contours for the majority of patients, a minority of patients require manual contouring and more in-depth physician review.

Uncertainty estimation and evaluation of deformation image registration based convolutional neural networks

Objective. Fast and accurate deformable image registration (DIR), including DIR uncertainty estimation, is essential for safe and reliable clinical deployment. While recent deep learning models have shown promise in predicting DIR with its uncertainty, challenges persist in proper uncertainty evaluation and hyperparameter optimization for these methods. This work aims to develop and evaluate a model that can perform fast DIR and predict its uncertainty in seconds. Approach. This study introduces a novel probabilistic multi-resolution image registration model utilizing convolutional neural networks to estimate a multivariate normal distributed dense displacement field (DDF) in a multimodal image registration problem. To assess the quality of the DDF distribution predicted by the model, we propose a new metric based on the Kullback–Leibler divergence. The performance of our approach was evaluated against three other DIR algorithms (VoxelMorph, Monte Carlo dropout, and Monte Carlo B-spline) capable of predicting uncertainty. The evaluation of the models included not only the quality of the deformation but also the reliability of the estimated uncertainty. Our application investigated the registration of a treatment planning computed tomography (CT) to follow-up cone beam CT for daily adaptive radiotherapy. Main results. The hyperparameter tuning of the models showed a trade-off between the estimated uncertainty’s reliability and the deformation’s accuracy. In the optimal trade-off, our model excelled in contour propagation and uncertainty estimation (p <0.05) compared to existing uncertainty estimation models. We obtained an average dice similarity coefficient of 0.89 and a KL-divergence of 0.15. Significance. By addressing challenges in DIR uncertainty estimation and evaluation, our work showed that both the DIR and its uncertainty can be reliably predicted, paving the way for safe deployment in a clinical environment.

Uncertainty estimation- and attention-based semi-supervised models for automatically delineate clinical target volume in CBCT images of breast cancer

ObjectivesAccurate segmentation of the clinical target volume (CTV) of CBCT images can observe the changes of CTV during patients' radiotherapy, and lay a foundation for the subsequent implementation of adaptive radiotherapy (ART). However, segmentation is challenging due to the poor quality of CBCT images and difficulty in obtaining target volumes. An uncertainty estimation- and attention-based semi-supervised model called residual convolutional block attention-uncertainty aware mean teacher (RCBA-UAMT) was proposed to delineate the CTV in cone-beam computed tomography (CBCT) images of breast cancer automatically.MethodsA total of 60 patients who undergone radiotherapy after breast-conserving surgery were enrolled in this study, which involved 60 planning CTs and 380 CBCTs. RCBA-UAMT was proposed by integrating residual and attention modules in the backbone network 3D UNet. The attention module can adjust channel and spatial weights of the extracted image features. The proposed design can train the model and segment CBCT images with a small amount of labeled data (5%, 10%, and 20%) and a large amount of unlabeled data. Four types of evaluation metrics, namely, dice similarity coefficient (DSC), Jaccard, average surface distance (ASD), and 95% Hausdorff distance (95HD), are used to assess the model segmentation performance quantitatively.ResultsThe proposed method achieved average DSC, Jaccard, 95HD, and ASD of 82%, 70%, 8.93, and 1.49 mm for CTV delineation on CBCT images of breast cancer, respectively. Compared with the three classical methods of mean teacher, uncertainty-aware mean-teacher and uncertainty rectified pyramid consistency, DSC and Jaccard increased by 7.89–9.33% and 14.75–16.67%, respectively, while 95HD and ASD decreased by 33.16–67.81% and 36.05–75.57%, respectively. The comparative experiment results of the labeled data with different proportions (5%, 10% and 20%) showed significant differences in the DSC, Jaccard, and 95HD evaluation indexes in the labeled data with 5% versus 10% and 5% versus 20%. Moreover, no significant differences were observed in the labeled data with 10% versus 20% among all evaluation indexes. Therefore, we can use only 10% labeled data to achieve the experimental objective.ConclusionsUsing the proposed RCBA-UAMT, the CTV of breast cancer CBCT images can be delineated reliably with a small amount of labeled data. These delineated images can be used to observe the changes in CTV and lay the foundation for the follow-up implementation of ART.

Deep-learning-based pelvic automatic segmentation in pelvic fractures

With the recent increase in traffic accidents, pelvic fractures are increasing, second only to skull fractures, in terms of mortality and risk of complications. Research is actively being conducted on the treatment of intra-abdominal bleeding, the primary cause of death related to pelvic fractures. Considerable preliminary research has also been performed on segmenting tumors and organs. However, studies on clinically useful algorithms for bone and pelvic segmentation, based on developed models, are limited. In this study, we explored the potential of deep-learning models presented in previous studies to accurately segment pelvic regions in X-ray images. Data were collected from X-ray images of 940 patients aged 18 or older at Gachon University Gil Hospital from January 2015 to December 2022. To segment the pelvis, Attention U-Net, Swin U-Net, and U-Net were trained, thereby comparing and analyzing the results using five-fold cross-validation. The Swin U-Net model displayed relatively high performance compared to Attention U-Net and U-Net models, achieving an average sensitivity, specificity, accuracy, and dice similarity coefficient of 96.77%, of 98.50%, 98.03%, and 96.32%, respectively.

Real-time coronary artery segmentation in CAG images: A semi-supervised deep learning strategy

BackgroundWhen treating patients with coronary artery disease and concurrent renal concerns, we often encounter a conundrum: how to achieve a clearer view of vascular details while minimizing the contrast and radiation doses during percutaneous coronary intervention (PCI). Our goal is to use deep learning (DL) to create a real-time roadmap for guiding PCI. To this end, segmentation, a critical first step, paves the way for detailed vascular analysis. Unlike traditional supervised learning, which demands extensive labeling time and manpower, our strategy leans toward semi-supervised learning. This method not only economizes on labeling efforts but also aims at reducing contrast and radiation exposure. Methods and resultsCAG data sourced from eight tertiary centers in Taiwan, comprising 500 labeled and 8952 unlabeled images. Employing 400 labels for training and reserving 100 for validation, we built a U-Net based network within a teacher-student architecture. The initial teacher model was updated with 8952 unlabeled images inputted, employing a quality control strategy involving consistency regularization and RandAugment. The optimized teacher model produced pseudo-labels for label expansion, which were then utilized to train the final student model. We attained an average dice similarity coefficient of 0.9003 for segmentation, outperforming supervised learning methods with the same label count. Even with only 5 % labels for semi-supervised training, the results surpassed a supervised method with 100 % labels inputted. This semi-supervised approach's advantage extends beyond single-frame prediction, yielding consistently superior results in continuous angiography films. ConclusionsHigh labeling cost hinders DL training. Semi-supervised learning, quality control, and pseudo-label expansion can overcome this. DL-assisted segmentation potentially provides a real-time PCI roadmap and further diminishes radiation and contrast doses.

Patient-specific temperature distribution prediction in laser interstitial thermal therapy: single-irradiation data-driven method

Laser interstitial thermal therapy (LITT) is popular for treating brain tumours and epilepsy. The strict control of tissue thermal damage extent is crucial for LITT. Temperature prediction is useful for predicting thermal damage extent. Accurately predicting in vivo brain tissue temperature is challenging due to the temperature dependence and the individual variations in tissue properties. Considering these factors is essential for improving the temperature prediction accuracy. Objective. To present a method for predicting patient-specific tissue temperature distribution within a target lesion area in the brain during LITT. Approach. A magnetic resonance temperature imaging (MRTI) data-driven estimation model was constructed and combined with a modified Pennes bioheat transfer equation (PBHE) to predict patient-specific temperature distribution. In the PBHE for temperature prediction, the individual specificity and temperature dependence of thermal tissue properties and blood perfusion, as well as the individual specificity of optical tissue properties were considered. Only MRTI data during one laser irradiation were required in the method. This enables the prediction of patient-specific temperature distribution and the resulting thermal damage region for subsequent ablations. Main results. Patient-specific temperature prediction was evaluated based on clinical data acquired during LITT in the brain, using intraoperative MRTI data as the reference standard. Our method significantly improved the prediction performance of temperature distribution and thermal damage region. The average root mean square error was decreased by 69.54%, the average intraclass correlation coefficient was increased by 37.5%, the average Dice similarity coefficient was increased by 43.14% for thermal damage region prediction. Significance. The proposed method can predict temperature distribution and thermal damage region at an individual patient level during LITT, providing a promising approach to assist in patient-specific treatment planning for LITT in the brain.

Deep learning-based efficient diagnosis of periapical diseases with dental X-rays

Diagnosis of periapical diseases usually requires a lot of time and effort from the dentist to perform a series of examinations and evaluations, although X-rays have been available to assist in visualizing the condition of the apical region of the tooth. To achieve efficient automated detection and diagnosis of periapical diseases (periapical granuloma, periapical abscess, periapical cysts, and condensing osteitis) by dental X-rays, a deep learning-based approach is proposed, which includes an improved Mask R-CNN-based lesion area segmentation network and a neural network classifier based on six texture features extracted from the segmented lesion regions. The results demonstrate that the average pixel accuracy of lesion area segmentation is over 0.97, while the average Dice Similarity Coefficient for segmentation is over 0.95. The accuracy, precision, recall, and F1 score of this method for the diagnosis of periapical diseases are all above 0.99. Precise segmentation and accurate classification of four distinct types of periapical diseases are achieved using this method, making it possible to create automated systems that can assist in the clinical diagnosis of periapical diseases, potentially streamlining the diagnostic process, reducing the workload of dentists, and improving overall patient care.

A Deep Learning Framework for Analysis of the Eustachian Tube and the Internal Carotid Artery.

Obtaining automated, objective 3-dimensional (3D) models of the Eustachian tube (ET) and the internal carotid artery (ICA) from computed tomography (CT) scans could provide useful navigational and diagnostic information for ET pathologies and interventions. We aim to develop a deep learning (DL) pipeline to automatically segment the ET and ICA and use these segmentations to compute distances between these structures. Retrospective cohort. Tertiary referral center. From a database of 30 CT scans, 60 ET and ICA pairs were manually segmented and used to train an nnU-Net model, a DL segmentation framework. These segmentations were also used to develop a quantitative tool to capture the magnitude and location of the minimum distance point (MDP) between ET and ICA. Performance metrics for the nnU-Net automated segmentations were calculated via the average Hausdorff distance (AHD) and dice similarity coefficient (DSC). The AHD for the ET and ICA were 0.922 and 0.246 mm, respectively. Similarly, the DSC values for the ET and ICA were 0.578 and 0.884. The mean MDP from ET to ICA in the cartilaginous region was 2.6 mm (0.7-5.3 mm) and was located on average 1.9 mm caudal from the bony cartilaginous junction. This study describes the first end-to-end DL pipeline for automated ET and ICA segmentation and analyzes distances between these structures. In addition to helping to ensure the safe selection of patients for ET dilation, this method can facilitate large-scale studies exploring the relationship between ET pathologies and the 3D shape of the ET.

Deep learning-aided 3D proxy-bridged region-growing framework for multi-organ segmentation

Accurate multi-organ segmentation in 3D CT images is imperative for enhancing computer-aided diagnosis and radiotherapy planning. However, current deep learning-based methods for 3D multi-organ segmentation face challenges such as the need for labor-intensive manual pixel-level annotations and high hardware resource demands, especially regarding GPU resources. To address these issues, we propose a 3D proxy-bridged region-growing framework specifically designed for the segmentation of the liver and spleen. Specifically, a key slice is selected from each 3D volume according to the corresponding intensity histogram. Subsequently, a deep learning model is employed to pinpoint the semantic central patch on this key slice, to calculate the growing seed. To counteract the impact of noise, segmentation of the liver and spleen is conducted on superpixel images created through proxy-bridging strategy. The segmentation process is then extended to adjacent slices by applying the same methodology iteratively, culminating in the comprehensive segmentation results. Experimental results demonstrate that the proposed framework accomplishes segmentation of the liver and spleen with an average Dice Similarity Coefficient of approximately 0.93 and a Jaccard Similarity Coefficient of around 0.88. These outcomes substantiate the framework's capability to achieve performance on par with that of deep learning methods, albeit requiring less guidance information and lower GPU resources.

Automatic segmentation of tumour and organs at risk in 3D MRI for cervical cancer radiation therapy with anatomical variations.

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.

A Hybrid Model for Fetal Growth Restriction Assessment by Automatic Placental Radiomics on T2-Weighted MRI and Multifeature Fusion.

MRI-based placental analyses have been used to improve fetal growth restriction (FGR) assessment by complementing ultrasound-based measurements. However, these are still limited by time-consuming manual annotation in MRI data and the lack of mother-based information. To develop and validate a hybrid model for accurate FGR assessment by automatic placental radiomics on T2-weighted imaging (T2WI) and multifeature fusion. Retrospective. 274 pregnant women (29.5 4.0 years) from two centers were included and randomly divided into training (N = 119), internal test (N = 40), time-independent validation (N = 43), and external validation (N = 72) sets. 1.5-T, T2WI half-Fourier acquisition single-shot turbo spin-echo pulse sequence. First, the placentas on T2WI were manually annotated, and a deep learning model was developed to automatically segment the placentas. Then, the radiomic features were extracted from the placentas and selected by three-step feature selection. In addition, fetus-based measurement features and mother-based clinical features were obtained from ultrasound examinations and medical records, respectively. Finally, a hybrid model based on random forest was constructed by fusing these features, and further compared with models based on other machine learning methods and different feature combinations. The performances of placenta segmentation and FGR assessment were evaluated by Dice similarity coefficient (DSC) and the area under the receiver operating characteristic curve (AUROC), respectively. A P-value <0.05 was considered statistically significant. The placentas were automatically segmented with an average DSC of 90.0%. The hybrid model achieved an AUROC of 0.923, 0.931, and 0.880 on the internal test, time-independent validation, and external validation sets, respectively. The mother-based clinical features resulted in significant performance improvements for FGR assessment. The proposed hybrid model may be able to assess FGR with high accuracy. Furthermore, information complementation based on placental, fetal, and maternal features could also lead to better FGR assessment performance. Stage 2.

A hybrid statistical morphometry free-form deformation approach to 3D personalized foot-ankle models

Foot and ankle joint models are widely used in the biomechanics community for musculoskeletal and finite element analysis. However, personalizing a foot and ankle joint model is highly time-consuming in terms of medical image collection and data processing. This study aims to develop and evaluate a framework for constructing a comprehensive 3D foot model that integrates statistical shape modeling (SSM) with free-form deformation (FFD) of internal bones. The SSM component is derived from external foot surface scans (skin measurements) of 50 participants, utilizing principal component analysis (PCA) to capture the variance in foot shapes. The derived surface shapes from SSM then guide the FFD process to accurately reconstruct the internal bone structures. The workflow accuracy was established by comparing three model-generated foot models against corresponding skin and bone geometries manually segmented and not part of the original training set. We used the top ten principal components representing 85 % of the population variation to create the model. For prediction validation, the average Dice similarity coefficient, Hausdorff distance error, and root mean square error were 0.92 ± 0.01, 2.2 ± 0.19 mm, and 2.95 ± 0.23 mm for soft tissues, and 0.84 ± 0.03, 1.83 ± 0.1 mm, and 2.36 ± 0.12 mm for bones, respectively. This study presents an efficient approach for 3D personalized foot model reconstruction via SSM generation of the foot surface that informs bone reconstruction based on FFD. The proposed workflow is part of the open-source Musculoskeletal Atlas Project linked to OpenSim and makes it feasible to accurately generate foot models informed by population anatomy, and suitable for rigid body analysis and finite element simulation.

Creating an AI-powered platform for neurosurgery alongside a usability examination: Progressing towards minimally invasive robotics

Recent advancements in artificial intelligence have paved the way for promising applications in neurosurgery, aiming to improve patient outcomes while minimizing risks. This paper introduces a novel AI-driven system designed to assist neurosurgeons in accurately identifying and localizing brain tumors. Leveraging deep learning algorithms, the system was trained on a comprehensive dataset of brain MRI scans for segmentation and classification tasks. Evaluation of the system on an independent set of brain MRI scans revealed an average Dice similarity coefficient of 0.87, indicating high performance. Moreover, a user experience assessment conducted at the Department of Neurosurgery, University Hospital Ulm, demonstrated notable enhancements in accuracy, efficiency, and reduced cognitive load and stress levels among users. Notably, the system showcased adaptability across various surgical scenarios and provided personalized guidance to users. These findings underscore the potential of AI to augment the quality of neurosurgical interventions and ultimately enhance patient outcomes. Future endeavors will focus on integrating this system with robotic surgical tools to facilitate minimally invasive surgeries.

Creating an AI-powered platform for neurosurgery alongside a usability examination: Progressing towards minimally invasive robotics

Recent advancements in artificial intelligence have paved the way for promising applications in neurosurgery, aiming to improve patient outcomes while minimizing risks. This paper introduces a novel AI-driven system designed to assist neurosurgeons in accurately identifying and localizing brain tumors. Leveraging deep learning algorithms, the system was trained on a comprehensive dataset of brain MRI scans for segmentation and classification tasks. Evaluation of the system on an independent set of brain MRI scans revealed an average Dice similarity coefficient of 0.87, indicating high performance. Moreover, a user experience assessment conducted at the Department of Neurosurgery, University Hospital Ulm, demonstrated notable enhancements in accuracy, efficiency, and reduced cognitive load and stress levels among users. Notably, the system showcased adaptability across various surgical scenarios and provided personalized guidance to users. These findings underscore the potential of AI to augment the quality of neurosurgical interventions and ultimately enhance patient outcomes. Future endeavors will focus on integrating this system with robotic surgical tools to facilitate minimally invasive surgeries.

Evaluation and mitigation of deformable image registration uncertainties for MRI-guided adaptive radiotherapy.

We evaluate the performance of a deformable image registration (DIR) software package in registering abdominal magnetic resonance images (MRIs) and then develop a mechanical modeling method to mitigate detected DIR uncertainties. Three evaluation metrics, namely mean displacement to agreement (MDA), DICE similarity coefficient (DSC), and standard deviation of Jacobian determinants (STD-JD), are used to assess the multi-modality (MM), contour-consistency (CC), and image-intensity (II)-based DIR algorithms in the MIM software package, as well as an in-house developed, contour matching-based finite element method (CM-FEM). Furthermore, we develop a hybrid FEM registration technique to modify the displacement vector field of each MIM registration. The MIM and FEM registrations were evaluated on MRIs obtained from 10 abdominal cancer patients. One-tailed Wilcoxon-Mann-Whitney (WMW) tests were conducted to compare the MIM registrations with their FEM modifications. For the registrations performed with the MIM-CC, MIM-MM, MIM-II, and CM-FEM algorithms, their average MDAs are 0.62±0.27, 2.39±1.30, 3.07±2.42, 1.04±0.72mm, and average DSCs are 0.94±0.03, 0.80±0.12, 0.77±0.15, 0.90±0.11, respectively. The p-values of the WMW tests between the MIM registrations and their FEM modifications are less than 0.0084 for STD-JDs and greater than 0.87 for MDA and DSC. Among the three MIM DIR algorithms, MIM-CC shows the smallest errors in terms of MDA and DSC but exhibits significant Jacobian uncertainties in the interior regions of abdominal organs. The hybrid FEM technique effectively mitigates the Jacobian uncertainties in these regions.

MRUNet-3D: A multi-stride residual 3D UNet for lung nodule segmentation

Obtaining an accurate segmentation of the pulmonary nodules in computed tomography (CT) images is challenging. This is due to: (1) the heterogeneous nature of the lung nodules; (2) comparable visual characteristics between the nodules and their surroundings. A robust multi-scale feature extraction mechanism that can effectively obtain multi-scale representations at a granular level can improve segmentation accuracy. As the most commonly used network in lung nodule segmentation, UNet, its variants, and other image segmentation methods lack this robust feature extraction mechanism. In this study, we propose a multi-stride residual 3D UNet (MRUNet-3D) to improve the segmentation accuracy of lung nodules in CT images. It incorporates a multi-slide Res2Net block (MSR), which replaces the simple sequence of convolution layers in each encoder stage to effectively extract multi-scale features at a granular level from different receptive fields and resolutions while conserving the strengths of 3D UNet. The proposed method has been extensively evaluated on the publicly available LUNA16 dataset. Experimental results show that it achieves competitive segmentation performance with an average dice similarity coefficient of 83.47 % and an average surface distance of 0.35 mm on the dataset. More notably, our method has proven to be robust to the heterogeneity of lung nodules. It has also proven to perform better at segmenting small lung nodules. Ablation studies have shown that the proposed MSR and RFIA modules are fundamental to improving the performance of the proposed model.

Evaluation of the consistency of the MRI- based AI segmentation cartilage model using the natural tibial plateau cartilage.

The study aims to evaluate the accuracy of an MRI-based artificial intelligence (AI) segmentation cartilage model by comparing it to the natural tibial plateau cartilage. This study included 33 patients (41 knees) with severe knee osteoarthritis scheduled to undergo total knee arthroplasty (TKA). All patients had a thin-section MRI before TKA. Our study is mainly divided into two parts: (i) In order to evaluate the MRI-based AI segmentation cartilage model's 2D accuracy, the natural tibial plateau was used as gold standard. The MRI-based AI segmentation cartilage model and the natural tibial plateau were represented in binary visualization (black and white) simulated photographed images by the application of Simulation Photography Technology. Both simulated photographed images were compared to evaluate the 2D Dice similarity coefficients (DSC). (ii) In order to evaluate the MRI-based AI segmentation cartilage model's 3D accuracy. Hand-crafted cartilage model based on knee CT was established. We used these hand-crafted CT-based knee cartilage model as gold standard to evaluate 2D and 3D consistency of between the MRI-based AI segmentation cartilage model and hand-crafted CT-based cartilage model. 3D registration technology was used for both models. Correlations between the MRI-based AI knee cartilage model and CT-based knee cartilage model were also assessed with the Pearson correlation coefficient. The AI segmentation cartilage model produced reasonably high two-dimensional DSC. The average 2D DSC between MRI-based AI cartilage model and the tibial plateau cartilage is 0.83. The average 2D DSC between the AI segmentation cartilage model and the CT-based cartilage model is 0.82. As for 3D consistency, the average 3D DSC between MRI-based AI cartilage model and CT-based cartilage model is 0.52. However, the quantification of cartilage segmentation with the AI and CT-based models showed excellent correlation (r = 0.725; P values < 0.05). Our study demonstrated that our MRI-based AI cartilage model can reliably extract morphologic features such as cartilage shape and defect location of the tibial plateau cartilage. This approach could potentially benefit clinical practices such as diagnosing osteoarthritis. However, in terms of cartilage thickness and three-dimensional accuracy, MRI-based AI cartilage model underestimate the actual cartilage volume. The previous AI verification methods may not be completely accurate and should be verified with natural cartilage images. Combining multiple verification methods will improve the accuracy of the AI model.

Combined use of radiomics and artificial neural networks for the three‐dimensional automatic segmentation of glioblastoma multiforme

AbstractGlioblastoma multiforme (GBM) is the most prevalent and aggressive primary brain tumour that has the worst prognosis in adults. Currently, the automatic segmentation of this kind of tumour is being intensively studied. Here, the automatic three‐dimensional segmentation of the GBM is achieved with its related subzones (active tumour, inner necrosis, and peripheral oedema). Preliminary segmentations were first defined based on the four basic magnetic resonance imaging modalities and classic image processing methods (multithreshold Otsu, Chan–Vese active contours, and morphological erosion). After an automatic gap‐filling post processing step, these preliminary segmentations were combined and corrected by a supervised artificial neural network of multilayer perceptron type with a hidden layer of 80 neurons, fed by 30 selected radiomic features of gray intensity and texture. Network classification has an overall accuracy of 83.9%, while the complete combined algorithm achieves average Dice similarity coefficients of 89.3%, 80.7%, 79.7%, and 66.4% for the entire region of interest, active tumour, oedema, and necrosis segmentations, respectively. These values are in the range of the best reported in the present bibliography, but even with better Hausdorff distances and lower computational costs. Results presented here evidence that it is possible to achieve the automatic segmentation of this kind of tumour by traditional radiomics. This has relevant clinical potential at the time of diagnosis, precision radiotherapy planning, or post‐treatment response evaluation.

Automated contouring, treatment planning, and quality assurance for VMAT craniospinal irradiation (VMAT-CSI).

Create a comprehensive automated solution for pediatric and adult VMAT-CSI including contouring, planning, and plan check to reduce planning time and improve plan quality. Seventy-seven previously treated CSI patients (age, 2-67 years) were used for creation of an auto-contouring model to segment 25 organs at risk (OARs). The auto-contoured OARs were evaluated using the Dice Similarity Coefficient (DSC), 95% Hausdorff Distance (HD95), and a qualitative ranking by one physician and one physicist (scale: 1-acceptable, 2-minor edits, 3-major edits). The auto-planning script was developed using the Varian Eclipse Scripting API and tested with 20 patients previously treated with either low-dose VMAT-CSI (12 Gy) or high-dose VMAT-CSI (36 Gy + 18 Gy boost). Clinically relevant metrics, planning time, and blinded physician review were used to evaluate significance of differences between the auto and manual plans. Finally, the plan preparation for treatment and plan check processes were automated to improve efficiency and safety of VMAT-CSI. The auto-contours achieved an average DSC of 0.71 ± 0.15, HD95 of 4.81 ± 4.68, and reviewers' ranking of 1.22 ± 0.39, indicating close to "acceptable-as-is" contours. Compared to the manual CSI plans, the auto-plans for both dose regimens achieved statistically significant reductions in body V50% and Dmean for parotids, submandibular, and thyroid glands. The variance in the dosimetric parameters decreased for the auto-plans as compared to the manual plans indicating better plan consistency. From the blinded review, the auto-plans were marked as equivalent or superior to the manual-plans 88.3% of the time. The required time for the auto-contouring and planning was consistently between 1-2 hours compared to an estimated 5-6 hours for manual contouring and planning. Reductions in contouring and planning time without sacrificing plan quality were obtained using the developed auto-planning process. The auto-planning scripts and documentation will be made freely available to other institutions and clinics.

Cascaded cross-attention transformers and convolutional neural networks for multi-organ segmentation in male pelvic computed tomography.

Segmentation of the prostate and surrounding organs at risk from computed tomography is required for radiation therapy treatment planning. We propose an automatic two-step deep learning-based segmentation pipeline that consists of an initial multi-organ segmentation network for organ localization followed by organ-specific fine segmentation. Initial segmentation of all target organs is performed using a hybrid convolutional-transformer model, axial cross-attention UNet. The output from this model allows for region of interest computation and is used to crop tightly around individual organs for organ-specific fine segmentation. Information from this network is also propagated to the fine segmentation stage through an image enhancement module, highlighting regions of interest in the original image that might be difficult to segment. Organ-specific fine segmentation is performed on these cropped and enhanced images to produce the final output segmentation. We apply the proposed approach to segment the prostate, bladder, rectum, seminal vesicles, and femoral heads from male pelvic computed tomography (CT). When tested on a held-out test set of 30 images, our two-step pipeline outperformed other deep learning-based multi-organ segmentation algorithms, achieving average dice similarity coefficient (DSC) of (prostate), (bladder), (rectum), (seminal vesicles), and (femoral heads). Our results demonstrate that a two-step segmentation pipeline with initial multi-organ segmentation and additional fine segmentation can delineate male pelvic CT organs well. The utility of this additional layer of fine segmentation is most noticeable in challenging cases, as our two-step pipeline produces noticeably more accurate and less erroneous results compared to other state-of-the-art methods on such images.