Median Dice Similarity Coefficient Research Articles

Quantitative texture analysis using machine learning for predicting interpretable pulmonary perfusion from non-contrast computed tomography in pulmonary embolism patients

BackgroundPulmonary embolism (PE) is life-threatening and requires timely and accurate diagnosis, yet current imaging methods, like computed tomography pulmonary angiography, present limitations, particularly for patients with contraindications to iodinated contrast agents. We aimed to develop a quantitative texture analysis pipeline using machine learning (ML) based on non-contrast thoracic computed tomography (CT) scans to discover intensity and textural features correlated with regional lung perfusion (Q) physiology and pathology and synthesize voxel-wise Q surrogates to assist in PE diagnosis.MethodsWe retrospectively collected 99mTc-labeled macroaggregated albumin Q-SPECT/CT scans from patients suspected of PE, including an internal dataset of 76 patients (64 for training, 12 for testing) and an external testing dataset of 49 patients. Quantitative CT features were extracted from segmented lung subregions and underwent a two-stage feature selection pipeline. The prior-knowledge-driven preselection stage screened for robust and non-redundant perfusion-correlated features, while the data-driven selection stage further filtered features by fitting ML models for classification. The final classification model, trained with the highest-performing PE-associated feature combination, was evaluated in the testing cohorts based on the Area Under the Curve (AUC) for subregion-level predictability. The voxel-wise Q surrogate was then synthesized using the final selected feature maps (FMs) and model score maps (MSMs) to investigate spatial distributions. The Spearman correlation coefficient (SCC) and Dice similarity coefficient (DSC) were used to assess the spatial consistency between FMs or MSMs and Q-SPECT scans.ResultsThe optimal model performance achieved an AUC of 0.863 during internal testing and 0.828 on the external testing cohort. The model identified a combination containing 14 intensity and textural features that were non-redundant, robust, and capable of distinguishing between high- and low-functional lung regions. Spatial consistency assessment in the internal testing cohort showed moderate-to-high agreement between MSMs and reference Q-SPECT scans, with median SCC of 0.66, median DSCs of 0.86 and 0.64 for high- and low-functional regions, respectively.ConclusionsThis study validated the feasibility of using quantitative texture analysis and a data-driven ML pipeline to generate voxel-wise lung perfusion surrogates, providing a radiation-free, widely accessible alternative to functional lung imaging in managing pulmonary vascular diseases.Clinical trial numberNot applicable.

Patient-specific deep learning tracking framework for real-time 2D target localization in MRI-guided radiotherapy

Purpose: We propose a tumor tracking framework for 2D cine MRI based on a pair of deep learning (DL) models relying on patient-specific (PS) training.Methods and materials: The chosen DL models are: 1) an image registration transformer and 2) an auto-segmentation convolutional neural network (CNN). We collected over 1,400,000 cine MRI frames from 219 patients treated on a 0.35 T MRI-linac plus 7,500 frames from additional 35 patients which were manually labelled and subdivided into fine-tuning, validation, and testing sets. The transformer was first trained on the unlabeled data (without segmentations). We then continued training (with segmentations) either on the fine-tuning set or for PS models based on eight randomly selected frames from the first 5 s of each patient's cine MRI. The PS auto-segmentation CNN was trained from scratch with the same eight frames for each patient, without pre-training. Furthermore, we implemented B-spline image registration as a conventional model, as well as different baselines. Output segmentations of all models were compared on the testing set using the Dice similarity coefficient (DSC), the 50% and 95% Hausdorff distance (HD50%/HD95%), and the root-mean-square-error of the target centroid in superior-inferior direction (RMSESI).Results: The PS transformer and CNN significantly outperformed all other models, achieving a median (inter-quartile range) DSC of 0.92 (0.03)/0.90 (0.04), HD50% of 1.0 (0.1)/1.0 (0.4) mm, HD95% of 3.1 (1.9)/3.8 (2.0) mm and RMSESI of 0.7 (0.4)/0.9 (1.0) mm on the testing set. Their inference time was about 36/8 ms per frame and PS fine-tuning required 3 min for labelling and 8/4 min for training. The transformer was better than the CNN in 9/12 patients, the CNN better in 1/12 patients and the two PS models achieved the same performance on the remaining 2/12 testing patients.Conclusion: For targets in the thorax, abdomen and pelvis, we found two PS DL models to provide accurate real-time target localization during MRI-guided radiotherapy.

Comprehensive segmentation of gray matter structures on T1-weighted brain MRI: A Comparative Study of CNN, CNN hybrid-transformer or -Mamba architectures.

Recent advances in deep learning have shown promising results in medical image analysis and segmentation. However, most brain MRI segmentation models are limited by the size of their datasets and/or the number of structures they can identify. This study evaluates the performance of six advanced deep learning models in segmenting 122 brain structures from T1-weighted MRI scans, aiming to identify the most effective model for clinical and research applications. 1,510 T1-weighted MRIs were used to compare six deep-learning models for the segmentation of 122 distinct gray matter structures: nnU-Net, SegResNet, SwinUNETR, UNETR, U-Mamba_BOT and U-Mamba_ Enc. Each model was rigorously tested for accuracy using the Dice Similarity Coefficient (DSC) and the 95th percentile Hausdorff Distance (HD95). Additionally, the volume of each structure was calculated and compared between normal control (NC) and Alzheimer's Disease (AD) patients. U-Mamba_Bot achieved the highest performance with a median DSC of 0.9112 [IQR:0.8957, 0.9250]. nnU-Net achieved a median DSC of 0.9027 [IQR: 0.8847, 0.9205] and had the highest HD95 of 1.392[IQR: 1.174, 2.029]. The value of each HD95 (<3mm) indicates its superior capability in capturing detailed brain structures accurately. Following segmentation, volume calculations were performed, and the resultant volumes of normal controls and AD patients were compared. The volume changes observed in thirteen brain substructures were all consistent with those reported in existing literature, reinforcing the reliability of the segmentation outputs. This study underscores the efficacy of U-Mamba_Bot as a robust tool for detailed brain structure segmentation in T1-weighted MRI scans. The congruence of our volumetric analysis with the literature further validates the potential of advanced deep-learning models to enhance the understanding of neurodegenerative diseases such as AD. Future research should consider larger datasets to validate these findings further and explore the applicability of these models in other neurological conditions. AD = Alzheimer's Disease; ADNI = Alzheimer's Disease Neuroimaging Initiative; DSC = Dice Similarity Coefficient; HD95 = the 95th Percentile Hausdorff Distance; IQR = Interquartile Range; NC = Normal Control; SSMs = State-space Sequence Models.

Minimizing human interference in an online fully automated daily adaptive radiotherapy workflow for bladder cancer

PurposeThe aim was to study the potential for an online fully automated daily adaptive radiotherapy (RT) workflow for bladder cancer, employing a focal boost and fiducial markers. The study focused on comparing the geometric and dosimetric aspects between the simulated automated online adaptive RT (oART) workflow and the clinically performed workflow.MethodsSeventeen patients with muscle-invasive bladder cancer were treated with daily Cone Beam CT (CBCT)-guided oART. The bladder and pelvic lymph nodes (CTVelective) received a total dose of 40 Gy in 20 fractions and the tumor bed received an additional simultaneously integrated boost (SIB) of 15 Gy (CTVboost). During the online sessions a CBCT was acquired and used as input for the AI-network to automatically delineate the bladder and rectum, i.e. influencers. These influencers were employed to guide the algorithm utilized in the delineation process of the target. Manual adjustments to the generated contours are common during this clinical workflow prior to plan reoptimization and RT delivery. To study the potential for an online fully automated workflow, the oART workflow was repeated in a simulation environment without manual adjustments. A comparison was made between the clinical and automatic contours and between the treatment plans optimized on these clinical (Dclin) and automatic contours (Dauto).ResultsThe bladder and rectum delineated by the AI-network differed from the clinical contours with a median Dice Similarity Coefficient of 0.99 and 0.92, a Mean Distance to Agreement of 1.9 mm and 1.3 mm and a relative volume of 100% and 95%, respectively. For the CTVboost these differences were larger, namely 0.71, 7 mm and 78%. For the CTVboost the median target coverage was 0.42% lower for Dauto compared to Dclin. For CTVelective this difference was 0.03%. The target coverage of Dauto met the clinical requirement of the CTV-coverage in 65% of the sessions for CTVboost and 95% of the sessions for the CTVelective.ConclusionsWhile an online fully automated daily adaptive RT workflow shows promise for bladder treatment, its complexity becomes apparent when incorporating a focal boost, necessitating manual checks to prevent potential underdosage of the target.

Development and comprehensive evaluation of a national DBCG consensus-based auto-segmentation model for lymph node levels in breast cancer radiotherapy

Background and purposeThis study aimed at training and validating a multi-institutional deep learning (DL) auto segmentation model for nodal clinical target volume (CTVn) in high-risk breast cancer (BC) patients with both training and validation dataset created with multi-institutional participation, with the overall aim of national clinical implementation in Denmark. Materials and methodsA gold standard (GS) dataset and a high-quality training dataset were created by 21 BC delineation experts from all radiotherapy centres in Denmark. The delineations were created according to ESTRO consensus delineation guidelines. Four models were trained: One per laterality and extension of CTVn internal mammary nodes. The DL models were tested quantitatively in their own test-set and in relation to interobserver variation (IOV) in the GS dataset with geometrical metrics, such as the Dice Similarity Coefficient (DSC). A blinded qualitative evaluation was conducted with a national board, presented to both DL and manual delineations. ResultsA median DSC > 0.7 was found for all, except the CTVn interpectoral node in one of the models. In the qualitative evaluation ‘no corrections needed’ were acquired for 297 (36 %) in the DL structures and 286 (34 %) for manual delineations. A higher rate of ‘major corrections’ and ‘easier to start from scratch’ was found in the manual delineations. The models performed within the IOV of an expert group, with two exceptions. ConclusionDL models were developed on a national consensus cohort and performed on par with the IOV between BC experts and had a comparable or higher clinical acceptance than expert manual delineations.

Convolutional Neural Networks for Segmentation of Pleural Mesothelioma: Analysis of Probability Map Thresholds (CALGB 30901, Alliance).

The purpose of this study was to evaluate the impact of probability map threshold on pleural mesothelioma (PM) tumor delineations generated using a convolutional neural network (CNN). One hundred eighty-six CT scans from 48 PM patients were segmented by a VGG16/U-Net CNN. A radiologist modified the contours generated at a 0.5 probability threshold. Percent difference of tumor volume and overlap using the Dice Similarity Coefficient (DSC) were compared between the reference standard provided by the radiologist and CNN outputs for thresholds ranging from 0.001 to 0.9. CNN-derived contours consistently yielded smaller tumor volumes than radiologist contours. Reducing the probability threshold from 0.5 to 0.01 decreased the absolute percent volume difference, on average, from 42.93% to 26.60%. Median and mean DSC ranged from 0.57 to 0.59, with a peak at a threshold of 0.2; no distinct threshold was found for percent volume difference. The CNN exhibited deficiencies with specific disease presentations, such as severe pleural effusion or disease in the pleural fissure. No single output threshold in the CNN probability maps was optimal for both tumor volume and DSC. This study emphasized the importance of considering both figures of merit when evaluating deep learning-based tumor segmentations across probability thresholds. This work underscores the need to simultaneously assess tumor volume and spatial overlap when evaluating CNN performance. While automated segmentations may yield comparable tumor volumes to that of the reference standard, the spatial region delineated by the CNN at a specific threshold is equally important.

Impact of geometric correction on echo-planar imaging-based apparent diffusion coefficient maps for abdominal radiotherapy

Objective. The apparent diffusion coefficient (ADC) extracted from diffusion-weighted magnetic resonance imaging (DWI) is a potential biomarker in radiotherapy (RT). DWI is often implemented with an echo-planar imaging (EPI) read-out due to speed, but unfortunately low geometric accuracy follows. This study aimed to investigate the influence of geometric distortions on the ADCs extracted from the gross tumor volume (GTV) and on the shape of the GTV in abdominal EPI-DWI. Approach. Twenty-one patients had EPI-DWI scans on a 1.5 T MRI sim before treatment and on a 1.5 T MRI-Linac at one of the first treatment fractions. Off-resonance correction with and without eddy current correction were applied to ADC maps. The clinical GTVs were deformed based on the same (but inverted) corrections to assess the local-regional geometric influence of distortions. Mean surface distance (MSD), Hausdorff distance (HD), and Dice similarity coefficient (DSC) were calculated to compare the original and distorted GTVs, and ADC values were calculated based on a mono-exponential model. Phantom measurements were performed to validate the applied correction method. Main results. The median (range) ADC change within the GTV after full distortion correction was 1.3% (0.02%–6.9%) for MRI-Sim and 1.5% (0.1%–6.4%) for MRI-Linac. The additional effect of the eddy current correction was small in both systems. The median (range) MSD, HD, and DSC comparing the original and off-resonance distorted GTVs for all patients were 0.43 mm (0.11–0.94 mm), 4.00 mm (1.00–7.81 mm) and 0.93 (0.82–0.99), respectively. Significance. Overall effect of distortion correction was small in terms of derived ADC values, indicating that distortion correction is unimportant for prediction of outcomes based on ADC. However, large local geometric changes occurred after off-resonance distortion correction for some patients, suggesting that if the spatial information from ADC maps is to be used for dose painting strategies, corrections should be applied.

Automated biventricular quantification in patients with repaired tetralogy of Fallot using a 3D deep learning segmentation model

BackgroundDeep learning is the state-of-the-art approach for automated segmentation of the left ventricle (LV) and right ventricle (RV) in cardiac magnetic resonance (CMR) images. However, these models have been mostly trained and validated using CMR datasets of structurally normal hearts or cases with acquired cardiac disease, and are therefore not well-suited to handle cases with congenital cardiac disease such as tetralogy of Fallot (TOF). We aimed to develop and validate a dedicated model with improved performance for LV and RV cavity and myocardium quantification in patients with repaired TOF. MethodsWe trained a 3D convolutional neural network (CNN) with 5-fold cross-validation using manually delineated end-diastolic (ED) and end-systolic (ES) short-axis image stacks obtained from either a public dataset containing patients with no or acquired cardiac pathology (n=100), an institutional dataset of TOF patients (n=96), or both datasets mixed. Our method allows for missing labels in the training images to accommodate for different ED and ES phases for LV and RV as is commonly the case in TOF. The best performing model was applied to all frames of a separate test set of TOF cases (n=36) and ED and ES phases were automatically determined for LV and RV separately. The model was evaluated against the performance of a commercial software (suiteHEART®, NeoSoft, Pewaukee, Wisconsin, US). ResultsTraining on the mixture of both datasets yielded the best agreement with the manual ground truth for the TOF cases, achieving a median DICE similarity coefficient of (93.8%, 89.8%) for LV cavity and of (92.9%, 90.9%) for RV cavity at (ED, ES) respectively, and of 80.9% and 61.8% for LV and RV myocardium at ED. The offset in automated ED and ES frame selection was 0.56 and 0.89 frames on average for LV and RV respectively. No statistically significant differences were found between our model and the commercial software for LV quantification (two-sided Wilcoxon signed rank test, p<5%), while RV quantification was significantly improved with our model achieving a mean absolute error of 12ml for RV cavity compared to 36ml for the commercial software. ConclusionWe developed and validated a fully automatic segmentation and quantification approach for LV and RV, including RV mass, in patients with repaired TOF. Compared to a commercial software, our approach is superior for RV quantification indicating its potential in clinical practice.

Deep learning with uncertainty estimation for automatic tumor segmentation in PET/CT of head and neck cancers: impact of model complexity, image processing and augmentation

Objective. Target volumes for radiotherapy are usually contoured manually, which can be time-consuming and prone to inter- and intra-observer variability. Automatic contouring by convolutional neural networks (CNN) can be fast and consistent but may produce unrealistic contours or miss relevant structures. We evaluate approaches for increasing the quality and assessing the uncertainty of CNN-generated contours of head and neck cancers with PET/CT as input. Approach. Two patient cohorts with head and neck squamous cell carcinoma and baseline 18F-fluorodeoxyglucose positron emission tomography and computed tomography images (FDG-PET/CT) were collected retrospectively from two centers. The union of manual contours of the gross primary tumor and involved nodes was used to train CNN models for generating automatic contours. The impact of image preprocessing, image augmentation, transfer learning and CNN complexity, architecture, and dimension (2D or 3D) on model performance and generalizability across centers was evaluated. A Monte Carlo dropout technique was used to quantify and visualize the uncertainty of the automatic contours. Main results. CNN models provided contours with good overlap with the manually contoured ground truth (median Dice Similarity Coefficient: 0.75–0.77), consistent with reported inter-observer variations and previous auto-contouring studies. Image augmentation and model dimension, rather than model complexity, architecture, or advanced image preprocessing, had the largest impact on model performance and cross-center generalizability. Transfer learning on a limited number of patients from a separate center increased model generalizability without decreasing model performance on the original training cohort. High model uncertainty was associated with false positive and false negative voxels as well as low Dice coefficients. Significance. High quality automatic contours can be obtained using deep learning architectures that are not overly complex. Uncertainty estimation of the predicted contours shows potential for highlighting regions of the contour requiring manual revision or flagging segmentations requiring manual inspection and intervention.

Pooled analysis of multiple sclerosis findings on multisite 7 Tesla MRI: Protocol and initial observations.

Although 7 T MRI research has contributed much to our understanding of multiple sclerosis (MS) pathology, most prior data has come from small, single-center studies with varying methods. In order to truly know if such findings have widespread applicability, multicenter methods and studies are needed. To address this, members of the North American Imaging in MS (NAIMS) Cooperative worked together to create a multicenter collaborative study of 7 T MRI in MS. In this manuscript, we describe the methods we have developed for the purpose of pooling together a large, retrospective dataset of 7 T MRIs acquired in multiple MS studies at five institutions. To date, this group has contributed five-hundred and twenty-eight 7 T MRI scans from 350 individuals with MS to a common data repository, with plans to continue to increase this sample size in the coming years. We have developed unified methods for image processing for data harmonization and lesion identification/segmentation. We report here our initial observations on intersite differences in acquisition, which includes site/device differences in brain coverage and image quality. We also report on the development of our methods and training of image evaluators, which resulted in median Dice Similarity Coefficients for trained raters' annotation of cortical and deep gray matter lesions, paramagnetic rim lesions, and meningeal enhancement between 0.73 and 0.82 compared to final consensus masks. We expect this publication to act as a resource for other investigators aiming to combine multicenter 7 T MRI datasets for the study of MS, in addition to providing a methodological reference for all future analysis projects to stem from the development of this dataset.

First implementation of artificial intelligence empowered all-in-one radiotherapy workflow for nasopharyngeal carcinoma.

134 Background: To establish an artificial intelligence (AI)-empowered All-In-One (AIO) workflow for providing one-stop radiotherapy to patients with nasopharyngeal carcinoma (NPC), and evaluate its feasibility and performance. Methods: The NPC-specific AIO workflow operates on a CT-integrated linac and integrates AI contouring and AI planning modules. A 3D U-net was employed for contours generation and 3D dose distribution. Knowledge-based planning approach and a protocol-based approach were used for plan optimization. MRI datasets with contours of primary gross tumors, CT datasets with contours of targets and organs at risk, and radiotherapy treatment planning datasets were collected from 877, 139, and 243 patients, respectively, for the training, validation, and testing of AI contouring and planning models. The NPC-specific AIO workflow underwent real-world clinical testing with an end-to-end evaluation involving 120 patients with NPC. The technical characteristics, implementation procedures and performance of AIO workflow were described and evaluated. Results: From March 2022 to October 2023, 120 patients with newly-diagnosed, non-metastatic NPC underwent AIO workflow, and 117 patients completed it in a median time of 23.2 min (range: 16.3 to 45.8 min). Median translation motions were 0.2 mm (CT scan to planning approval) and 0.1 mm (during beam delivery). AI-generated contours required little revision for clinical target volume (CTV1) and organs at risk (OARs), minor revision for cervical lymph nodes (GTVn) and clinical target volume (CTV2), with median Dice similarity coefficients (DSCs) of 0.98 and 0.94, respectively; more revision for GTVp, with a median DSC of 0.84. Of the 117 AI-generated plannings, 108 (92.3%) passed after initial auto-optimization. Volumes receiving ≥ 100% of the prescribed dose for all planning target volumes was ≥ 97.8%. Mean dosimetric parameters met dose constraints for all OARs except the thyroid and submandibulars. One patient died from acute renal failure after 10 fractions of radiotherapy. At week 12 after radiotherapy completion, 115 patients achieved complete response;1 had stable disease. Grade 3-4 acute effects were reported by 54 of 115 patients (36%). Conclusions: We successfully established an AI-empowered AIO workflow for initial radiotherapy of NPC, confirming its feasibility and safety for clinical application.

Validation of an automated segmentation method for body composition analysis in colorectal cancer patients using diagnostic abdominal computed tomography images

Background & AimsSeveral automated programs have been developed to facilitate body composition analysis of images from abdominal computed tomography (CT) scans. External validation in patients with colorectal cancer is necessary for use in research and clinical practice. Our aim was to validate an automatic method (AutoMATiCA) of segmenting CT images at the third lumbar level (L3) from patients with colorectal cancer, by comparing with manual segmentation. MethodsDiagnostic abdominal CT scans of consecutive patients with stage I-III colorectal cancer were analysed to measure cross-sectional areas and tissue densities of skeletal muscle and intra-muscular, visceral, and subcutaneous adipose tissue. Trained analysts performed manual segmentation of L3 CT images using SliceOmatic. Automatic segmentation was performed using AutoMATiCA, an open-source software. The Dice similarity coefficient (DSC) was calculated to assess segmentation accuracy. Agreement of automatic with manual segmentation was evaluated using intra-class correlation coefficients (ICCs) and Bland-Altman plots with limits of agreement. ResultsA total of 292 scans were included, of which 62% were from male patients. The agreement of AutoMATiCA with the manual segmentation was excellent, with median DSC values ranging from 0.900-0.991 and ICCs above 0.95 for all segmented areas. No systematic deviations were observed in Bland-Altman plots for all segmented areas, with overall narrow limits of agreement. ConclusionsAutoMATiCA provides an accurate segmentation of abdominal CT images from patients with colorectal cancer. Our findings support its use as a highly efficient automated tool for body composition analysis in research and potentially also in clinical practice.

Improving 3D dose prediction for breast radiotherapy using novel glowing masks and gradient-weighted loss functions.

The quality of treatment plans for breast cancer can vary greatly. This variation could be reduced by using dose prediction to automate treatment planning. Our work investigates novel methods for training deep-learning models that are capable of producing high-quality dose predictions for breast cancer treatment planning. The goal of this work was to compare the performance impact of two novel techniques for deep learning dose prediction models for tangent field treatments for breast cancer. The first technique, a "glowing" mask algorithm, encodes the distance from a contour into each voxel in a mask. The second, a gradient-weighted mean squared error (MSE) loss function, emphasizes the error in high-dose gradient regions in the predicted image. Four 3D U-Net deep learning models were trained using the planning CT and contours of the heart, lung, and tumor bed as inputs. The dataset consisted of 305 treatment plans split into 213/46/46 training/validation/test sets using a 70/15/15% split. We compared the impact of novel "glowing" anatomical mask inputs and a novel gradient-weighted MSE loss function to their standard counterparts, binary anatomical masks, and MSE loss, using an ablation study methodology. To assess performance, we examined the mean error and mean absolute error (ME/MAE) in dose across all within-body voxels, the error in mean dose to heart, ipsilateral lung, and tumor bed, dice similarity coefficient (DSC) across isodose volumes defined by 0%-100% prescribed dose thresholds, and gamma analysis (3%/3mm). The combination of novel glowing masks and gradient weighted loss function yielded the best-performing model in this study. This model resulted in a mean ME of 0.40%, MAE of 2.70%, an error in mean dose to heart and lung of -0.10 and 0.01Gy, and an error in mean dose to the tumor bed of -0.01%. The median DSC at 50/95/100% isodose levels were 0.91/0.87/0.82. The mean 3D gamma pass rate (3%/3mm) was 93%. This study found the combination of novel anatomical mask inputs and loss function for dose prediction resulted in superior performance to their standard counterparts. These results have important implications for the field of radiotherapy dose prediction, as the methods used here can be easily incorporated into many other dose prediction models for other treatment sites. Additionally, this dose prediction model for breast radiotherapy has sufficient performance to be used in an automated planning pipeline for tangent field radiotherapy and has the major benefit of not requiring a PTV for accurate dose prediction.

Convolutional Neural Network-Based Automated Segmentation of Skeletal Muscle and Subcutaneous Adipose Tissue on Thigh MRI in Muscular Dystrophy Patients.

We aim to develop a deep learning-based algorithm for automated segmentation of thigh muscles and subcutaneous adipose tissue (SAT) from T1-weighted muscle MRIs from patients affected by muscular dystrophies (MDs). From March 2019 to February 2022, adult and pediatric patients affected by MDs were enrolled from Azienda Ospedaliera Universitaria Pisana, Pisa, Italy (Institution 1) and the IRCCS Stella Maris Foundation, Calambrone-Pisa, Italy (Institution 2), respectively. All patients underwent a bilateral thighs MRI including an axial T1 weighted in- and out-of-phase (dual-echo). Both muscles and SAT were manually and separately segmented on out-of-phase image sets by a radiologist with 6 years of experience in musculoskeletal imaging. A U-Net1 and U-Net3 were built to automatically segment the SAT, all the thigh muscles together and the three muscular compartments separately. The dataset was randomly split into the on train, validation, and test set. The segmentation performance was assessed through the Dice similarity coefficient (DSC). The final cohort included 23 patients. The estimated DSC for U-Net1 was 96.8%, 95.3%, and 95.6% on train, validation, and test set, respectively, while the estimated accuracy for U-Net3 was 94.1%, 92.9%, and 93.9%. Both of the U-Nets achieved a median DSC of 0.95 for SAT segmentation. The U-Net1 and the U-Net3 achieved an optimal agreement with manual segmentation for the automatic segmentation. The so-developed neural networks have the potential to automatically segment thigh muscles and SAT in patients affected by MDs.

Automated segmentation of brain metastases with deep learning: A multi-center, randomized crossover, multi-reader evaluation study.

Artificial intelligence has been proposed for brain metastasis (BM) segmentation but it has not been fully clinically validated. The aim of this study was to develop and evaluate a system for BM segmentation. A deep-learning-based BM segmentation system (BMSS) was developed using contrast-enhanced MR images from 488 patients with 10338 brain metastases. A randomized crossover, multi-reader study was then conducted to evaluate the performance of the BMSS for BM segmentation using data prospectively collected from 50 patients with 203 metastases at 5 centers. Five radiology residents and 5 attending radiologists were randomly assigned to contour the same prospective set in assisted and unassisted modes. Aided and unaided Dice similarity coefficients (DSCs) and contouring times per lesion were compared. The BMSS alone yielded a median DSC of 0.91 (95% confidence interval, 0.90-0.92) in the multi-center set and showed comparable performance between the internal and external sets (P = .67). With BMSS assistance, the readers increased the median DSC from 0.87 (0.87-0.88) to 0.92 (0.92-0.92) (P < .001) with a median time saving of 42% (40-45%) per lesion. Resident readers showed a greater improvement than attending readers in contouring accuracy (improved median DSC, 0.05 [0.05-0.05] vs 0.03 [0.03-0.03]; P < .001), but a similar time reduction (reduced median time, 44% [40-47%] vs 40% [37-44%]; P = .92) with BMSS assistance. The BMSS can be optimally applied to improve the efficiency of brain metastasis delineation in clinical practice.

Stepwise Transfer Learning for Expert-level Pediatric Brain Tumor MRI Segmentation in a Limited Data Scenario.

Purpose To develop, externally test, and evaluate clinical acceptability of a deep learning pediatric brain tumor segmentation model using stepwise transfer learning. Materials and Methods In this retrospective study, the authors leveraged two T2-weighted MRI datasets (May 2001 through December 2015) from a national brain tumor consortium (n = 184; median age, 7 years [range, 1-23 years]; 94 male patients) and a pediatric cancer center (n = 100; median age, 8 years [range, 1-19 years]; 47 male patients) to develop and evaluate deep learning neural networks for pediatric low-grade glioma segmentation using a stepwise transfer learning approach to maximize performance in a limited data scenario. The best model was externally tested on an independent test set and subjected to randomized blinded evaluation by three clinicians, wherein they assessed clinical acceptability of expert- and artificial intelligence (AI)-generated segmentations via 10-point Likert scales and Turing tests. Results The best AI model used in-domain stepwise transfer learning (median Dice score coefficient, 0.88 [IQR, 0.72-0.91] vs 0.812 [IQR, 0.56-0.89] for baseline model; P = .049). With external testing, the AI model yielded excellent accuracy using reference standards from three clinical experts (median Dice similarity coefficients: expert 1, 0.83 [IQR, 0.75-0.90]; expert 2, 0.81 [IQR, 0.70-0.89]; expert 3, 0.81 [IQR, 0.68-0.88]; mean accuracy, 0.82). For clinical benchmarking (n = 100 scans), experts rated AI-based segmentations higher on average compared with other experts (median Likert score, 9 [IQR, 7-9] vs 7 [IQR 7-9]) and rated more AI segmentations as clinically acceptable (80.2% vs 65.4%). Experts correctly predicted the origin of AI segmentations in an average of 26.0% of cases. Conclusion Stepwise transfer learning enabled expert-level automated pediatric brain tumor autosegmentation and volumetric measurement with a high level of clinical acceptability. Keywords: Stepwise Transfer Learning, Pediatric Brain Tumors, MRI Segmentation, Deep Learning Supplemental material is available for this article. © RSNA, 2024.

Navigated intraoperative ultrasound in pediatric brain tumors

PurposeThe aim of this study was to evaluate the diagnostic value and accuracy of navigated intraoperative ultrasound (iUS) in pediatric oncological neurosurgery as compared to intraoperative magnetic resonance imaging (iMRI).MethodsA total of 24 pediatric patients undergoing tumor debulking surgery with iUS, iMRI, and neuronavigation were included in this study. Prospective acquisition of iUS images was done at two time points during the surgical procedure: (1) before resection for tumor visualization and (2) after resection for residual tumor assessment. Dice similarity coefficients (DSC), Hausdorff distances 95th percentiles (HD95) and volume differences, sensitivity, and specificity were calculated for iUS segmentations as compared to iMRI.ResultsA high correlation (R = 0.99) was found for volume estimation as measured on iUS and iMRI before resection. A good spatial accuracy was demonstrated with a median DSC of 0.72 (IQR 0.14) and a median HD95 percentile of 4.98 mm (IQR 2.22 mm). The assessment after resection demonstrated a sensitivity of 100% and a specificity of 84.6% for residual tumor detection with navigated iUS. A moderate accuracy was observed with a median DSC of 0.58 (IQR 0.27) and a median HD95 of 5.84 mm (IQR 4.04 mm) for residual tumor volumes.ConclusionWe found that iUS measurements of tumor volume before resection correlate well with those obtained from preoperative MRI. The accuracy of residual tumor detection was reliable as compared to iMRI, indicating the suitability of iUS for directing the surgeon’s attention to areas suspect for residual tumor. Therefore, iUS is considered as a valuable addition to the neurosurgical armamentarium.Trial registration number and datePMCLAB2023.476, February 12th 2024.

Development and benchmarking of a Deep Learning-based MRI-guided gross tumor segmentation algorithm for Radiomics analyses in extremity soft tissue sarcomas

BackgroundVolume of interest (VOI) segmentation is a crucial step for Radiomics analyses and radiotherapy (RT) treatment planning. Because it can be time-consuming and subject to inter-observer variability, we developed and tested a Deep Learning-based automatic segmentation (DLBAS) algorithm to reproducibly predict the primary gross tumor as VOI for Radiomics analyses in extremity soft tissue sarcomas (STS). MethodsA DLBAS algorithm was trained on a cohort of 157 patients and externally tested on an independent cohort of 87 patients using contrast-enhanced MRI. Manual tumor delineations by a radiation oncologist served as ground truths (GTs). A benchmark study with 20 cases from the test cohort compared the DLBAS predictions against manual VOI segmentations of two residents (ERs) and clinical delineations of two radiation oncologists (ROs). The ROs rated DLBAS predictions regarding their direct applicability. ResultsThe DLBAS achieved a median dice similarity coefficient (DSC) of 0.88 against the GTs in the entire test cohort (interquartile range (IQR): 0.11) and a median DSC of 0.89 (IQR 0.07) and 0.82 (IQR 0.10) in comparison to ERs and ROs, respectively. Radiomics feature stability was high with a median intraclass correlation coefficient of 0.97, 0.95 and 0.94 for GTs, ERs, and ROs, respectively. DLBAS predictions were deemed clinically suitable by the two ROs in 35% and 20% of cases, respectively. ConclusionThe results demonstrate that the DLBAS algorithm provides reproducible VOI predictions for radiomics feature extraction. Variability remains regarding direct clinical applicability of predictions for RT treatment planning.

Deep learning for automatic segmentation of vestibular schwannoma: a retrospective study from multi-center routine MRI.

Automatic segmentation of vestibular schwannoma (VS) from routine clinical MRI has potential to improve clinical workflow, facilitate treatment decisions, and assist patient management. Previous work demonstrated reliable automatic segmentation performance on datasets of standardized MRI images acquired for stereotactic surgery planning. However, diagnostic clinical datasets are generally more diverse and pose a larger challenge to automatic segmentation algorithms, especially when post-operative images are included. In this work, we show for the first time that automatic segmentation of VS on routine MRI datasets is also possible with high accuracy. We acquired and publicly release a curated multi-center routine clinical (MC-RC) dataset of 160 patients with a single sporadic VS. For each patient up to three longitudinal MRI exams with contrast-enhanced T1-weighted (ceT1w) (n = 124) and T2-weighted (T2w) (n = 363) images were included and the VS manually annotated. Segmentations were produced and verified in an iterative process: (1) initial segmentations by a specialized company; (2) review by one of three trained radiologists; and (3) validation by an expert team. Inter- and intra-observer reliability experiments were performed on a subset of the dataset. A state-of-the-art deep learning framework was used to train segmentation models for VS. Model performance was evaluated on a MC-RC hold-out testing set, another public VS datasets, and a partially public dataset. The generalizability and robustness of the VS deep learning segmentation models increased significantly when trained on the MC-RC dataset. Dice similarity coefficients (DSC) achieved by our model are comparable to those achieved by trained radiologists in the inter-observer experiment. On the MC-RC testing set, median DSCs were 86.2(9.5) for ceT1w, 89.4(7.0) for T2w, and 86.4(8.6) for combined ceT1w+T2w input images. On another public dataset acquired for Gamma Knife stereotactic radiosurgery our model achieved median DSCs of 95.3(2.9), 92.8(3.8), and 95.5(3.3), respectively. In contrast, models trained on the Gamma Knife dataset did not generalize well as illustrated by significant underperformance on the MC-RC routine MRI dataset, highlighting the importance of data variability in the development of robust VS segmentation models. The MC-RC dataset and all trained deep learning models were made available online.

Design and evaluation of a deep learning-based automatic segmentation of maxillary and mandibular substructures using a 3D U-Net

BackgroundCurrent segmentation approaches for radiation treatment planning in head and neck cancer patients (HNCP) typically consider the entire mandible as an organ at risk, whereas segmentation of the maxilla remains uncommon. Accurate risk assessment for osteoradionecrosis (ORN) or implant-based dental rehabilitation after radiation therapy may require a nuanced analysis of dose distribution in specific mandibular and maxillary segments. Manual segmentation is time-consuming and inconsistent, and there is no definition of jaw subsections. Materials and methodsThe mandible and maxilla were divided into 12 substructures. The model was developed from 82 computed tomography (CT) scans of HNCP and adopts an encoder-decoder three-dimensional (3D) U-Net structure. The efficiency and accuracy of the automated method were compared against manual segmentation on an additional set of 20 independent CT scans. The evaluation metrics used were the Dice similarity coefficient (DSC), 95% Hausdorff distance (HD95), and surface DSC (sDSC). ResultsAutomated segmentations were performed in a median of 86 s, compared to manual segmentations, which took a median of 53.5 min. The median DSC per substructure ranged from 0.81 to 0.91, and the median HD95 ranged from 1.61 to 4.22. The number of artifacts did not affect these scores. The maxillary substructures showed lower metrics than the mandibular substructures. ConclusionsThe jaw substructure segmentation demonstrated high accuracy, time efficiency, and promising results in CT scans with and without metal artifacts. This novel model could provide further investigation into dose relationships with ORN or dental implant failure in normal tissue complication prediction models.