Mean Dice Similarity Coefficient Research Articles

Deep Learning-Based Precontrast CT Parcellation for MRI-Free Brain Amyloid PET Quantification.

This study aimed to develop a deep learning (DL) model for brain region parcellation using CT data from PET/CT scans to enable accurate amyloid quantification in 18F-FBB PET/CT without relying on high-resolution MRI. A retrospective dataset of PET/CT and T1-weighted MRI pairs from 226 individuals (157 with mild cognitive impairment or dementia and 69 healthy controls) was used. The dataset was split into training/validation (60%) and test (40%) sets. Utilizing auto-generated segmentation labels, 3 UNets were independently trained for multiplanar brain parcellation on CT and subsequently ensembled. Amyloid load was measured across 46 volumes of interest (VOIs), derived from the Desikan-Killiany-Tourville atlas. Dice similarity coefficient between the proposed CT-based DL model and MRI-based (FreeSurfer) method was calculated, with SUVR comparison using linear regression analysis and intraclass correlation coefficient. Global SUVRs were also compared within groups with clinical dementia ratings (CDRs) of 0, 0.5, and 1. The DL-based CT parcellation achieved mean Dice similarity coefficients of 0.80 for all 46 VOIs, 0.72 for 16 cortical and limbic VOIs, and 0.83 for 30 subcortical VOIs. For regional and global SUVR comparisons, the linear regression yielded a slope, y-intercept, and R2 of 1 ± 0.027, 0 ± 0.040, and ≧0.976, respectively (P < 0.001), and the intraclass correlation coefficient was ≧0.988 (P < 0.001). For global SUVRs in each CDR group, these values were 1 ± 0.020, 0 ± 0.026, ≧0.993, and ≧0.996, respectively (P < 0.001). Both MRI-based and CT-based global SUVR showed a consistent increase as the CDR score increased. The DL-based CT parcellation agrees strongly with MRI-based methods for amyloid PET quantification.

Combining Multifrequency Magnetic Resonance Elastography With Automatic Segmentation to Assess Renal Function in Patients With Chronic Kidney Disease.

Multifrequency MR elastography (mMRE) enables noninvasive quantification of renal stiffness in patients with chronic kidney disease (CKD). Manual segmentation of the kidneys on mMRE is time-consuming and prone to increased interobserver variability. To evaluate the performance of mMRE combined with automatic segmentation in assessing CKD severity. Prospective. A total of 179 participants consisting of 95 healthy volunteers and 84 participants with CKD. 3 T, single shot spin echo planar imaging sequence. Participants were randomly assigned into training (n = 58), validation (n = 15), and test (n = 106) sets. Test set included 47 healthy volunteers and 58 CKD participants with different stages (21 stage 1-2, 22 stage 3, and 16 stage 4-5) based on estimated glomerular filtration rate (eGFR). Shear wave speed (SWS) values from mMRE was measured using automatic segmentation constructed through the nnU-Net deep-learning network. Standard manual segmentation was created by a radiologist. In the test set, the automatically segmented renal SWS were compared between healthy volunteers and CKD subgroups, with age as a covariate. The association between SWS and eGFR was investigated in participants with CKD. Dice similarity coefficient (DSC), analysis of covariance, Pearson and Spearman correlation analyses. P < 0.05 was considered statistically significant. Mean DSCs between standard manual and automatic segmentation were 0.943, 0.901, and 0.970 for the renal cortex, medulla, and parenchyma, respectively. The automatically quantified cortical, medullary, and parenchymal SWS were significantly correlated with eGFR (r = 0.620, 0.605, and 0.640, respectively). Participants with CKD stage 1-2 exhibited significantly lower cortical SWS values compared to healthy volunteers (2.44 ± 0.16 m/second vs. 2.56 ± 0.17 m/second), after adjusting age. mMRE combined with automatic segmentation revealed abnormal renal stiffness in patients with CKD, even with mild renal impairment. The renal stiffness of patients with chronic kidney disease varies according to the function and structure of the kidney. This study integrates multifrequency magnetic resonance elastography with automated segmentation technique to assess renal stiffness in patients with chronic kidney disease. The findings indicate that this method is capable of distinguishing between patients with chronic kidney disease, including those with mild renal impairment, while simultaneously reducing the subjectivity and time required for radiologists to analyze images. This research enhances the efficiency of image processing for radiologists and assists nephrologists in detecting early-stage damage in patients with chronic kidney disease. 2 TECHNICAL EFFICACY: Stage 2.

Evaluation of Siemens Healthineers’ StrokeSegApp for automated diffusion and perfusion lesion segmentation in patients with ischemic stroke

PurposeThis study aimed to evaluate the perfomance of Siemens Healthineers’ StrokeSegApp performance in automatically segmenting diffusion and perfusion lesions in patients with acute ischemic stroke and to assess its clinical utility in guiding mechanical thrombectomy decisions.MethodsThis retrospective study used MRI data of acute ischemic stroke patients from the prospective observational single-center 1000Plus study, acquired between September 2008 and June 2013 (clinicaltrials.org; NCT00715533) and manually segmented by radiologists as the ground truth. The performance of the StrokeSegApp was compared against this ground truth using the dice similarity coefficient (DSC) and Bland–Altman plots. The study also evaluated the application’s ability to recommend mechanical thrombectomy based on DEFUSE 2 and 3 trial criteria.ResultsThe StrokeSegApp demonstrated a mean DSC of 0.60 (95% CI: 0.57–0.63; n = 241) for diffusion deficit segmentation and 0.80 (95% CI: 0.76–0.85; n = 56) for perfusion deficit segmentation. The mean volume deviation was 0.49 mL for diffusion lesions and −7.69 mL for perfusion lesions. Out of 56 subjects meeting DEFUSE 2/3 criteria in the cohort, it correctly identified mechanical thrombectomy candidates with a sensitivity of 82.1% (95% CI: 63.1–93.9%) and a specificity of 96.4% (95% CI: 81.7–99.9%).ConclusionThe Siemens Healthineers’ StrokeSegApp provides accurate automated segmentation of ischemic stroke lesions, comparable to human experts as well as similar commercial software, and shows potential as a reliable tool in clinical decision-making for stroke treatment.

Deep learning-based synthetic CT for dosimetric monitoring of combined conventional radiotherapy and lattice boost in large lung tumors

PurposeConventional radiotherapy (CRT) has limited local control and poses a high risk of severe toxicity in large lung tumors. This study aimed to develop an integrated treatment plan that combines CRT with lattice boost radiotherapy (LRT) and monitors its dosimetric characteristics.MethodsThis study employed cone-beam computed tomography from 115 lung cancer patients to develop a U-Net + + deep learning model for generating synthetic CT (sCT). The clinical feasibility of sCT was thoroughly evaluated in terms of image clarity, Hounsfield Unit (HU) consistency, and computational accuracy. For large lung tumors, accumulated doses to the gross tumor volume (GTV) and organs at risk (OARs) during 20 fractions of CRT were precisely monitored using matrices derived from the deformable registration of sCT and planning CT (pCT). Additionally, for patients with minimal tumor shrinkage during CRT, an sCT-based adaptive LRT boost plan was introduced, with its dosimetric properties, treatment safety in high dose regions, and delivery accuracy quantitatively assessed.ResultsThe image quality and HU consistency of sCT improved significantly, with dose deviations ranging from 0.15% to 1.25%. These results indicated that sCT is feasible for inter-fraction dose monitoring and adaptive planning. After rigid and hybrid deformable registration of sCT and pCT, the mean distance-to-agreement was 0.80 ± 0.18 mm, and the mean Dice similarity coefficient was 0.97 ± 0.01. Monitoring dose accumulation over 20 CRT fractions showed an increase in high-dose regions of the GTV (P < 0.05) and a reduction in low-dose regions (P < 0.05). Dosimetric parameters of all OARs were significantly higher than those in the original treatment plan (P < 0.01). The sCT based adaptive LRT boost plan, when combined with CRT, significantly reduced the dose to OARs compared to CRT alone (P < 0.05). In LRT plan, high-dose regions for the GTV and D95% exhibited displacements greater than 5 mm from the tumor boundary in 19 randomly scanned sCT sequences under free breathing conditions. Validation of dose delivery using TLD phantom measurements showed that more than half of the dose points in the sCT based LRT plan had deviations below 2%, with a maximum deviation of 5.89%.ConclusionsThe sCT generated by the U-Net + + model enhanced the accuracy of monitoring the actual accumulated dose, thereby facilitating the evaluation of therapeutic efficacy and toxicity. Additionally, the sCT-based LRT boost plan, combined with CRT, further minimized the dose delivered to OARs while ensuring safe and precise treatment delivery.

Enhanced accuracy and stability in automated intra-pancreatic fat deposition monitoring of type 2 diabetes mellitus using Dixon MRI and deep learning.

Intra-pancreatic fat deposition (IPFD) is closely associated with the onset and progression of type 2 diabetes mellitus (T2DM). We aimed to develop an accurate and automated method for assessing IPFD on multi-echo Dixon MRI. In this retrospective study, 534 patients from two centers who underwent upper abdomen MRI and completed multi-echo and double-echo Dixon MRI were included. A pancreatic segmentation model was trained on double-echo Dixon water images using nnU-Net. Predicted masks were registered to the proton density fat fraction (PDFF) maps of the multi-echo Dixon sequence. Deep semantic segmentation feature-based radiomics (DSFR) and radiomics features were separately extracted on the PDFF maps and modeled using the support vector machine method with 5-fold cross-validation. The first deep learning radiomics (DLR) model was constructed to distinguish T2DM from non-diabetes and pre-diabetes by averaging the output scores of the DSFR and radiomics models. The second DLR model was then developed to distinguish pre-diabetes from non-diabetes. Two radiologist models were constructed based on the mean PDFF of three pancreatic regions of interest. The mean Dice similarity coefficient for pancreas segmentation was 0.958 in the total test cohort. The AUCs of the DLR and two radiologist models in distinguishing T2DM from non-diabetes and pre-diabetes were 0.868, 0.760, and 0.782 in the training cohort, and 0.741, 0.724, and 0.653 in the external test cohort, respectively. For distinguishing pre-diabetes from non-diabetes, the AUCs were 0.881, 0.688, and 0.688 in the training cohort, which included data combined from both centers. Testing was not conducted due to limited pre-diabetic patients. Intraclass correlation coefficients between radiologists' pancreatic PDFF measurements were 0.800 and 0.699 at two centers, suggesting good and moderate reproducibility, respectively. The DLR model demonstrated superior performance over radiologists, providing a more efficient, accurate and stable method for monitoring IPFD and predicting the risk of T2DM and pre-diabetes. This enables IPFD assessment to potentially serve as an early biomarker for T2DM, providing richer clinical information for disease progression and management.

Large language model-augmented learning for auto-delineation of treatment targets in head-and-neck cancer radiotherapy.

Radiation therapy (RT) is highly effective, but its success depends on accurate, manual target delineation, which is time-consuming, labor-intensive, and prone to variability. Despite AI advancements in auto-contouring normal tissues, accurate RT target volume delineation remains challenging. This study presents Radformer, a novel visual language model that integrates text-rich clinical data with medical imaging for accurate automated RT target volume delineation. We developed Radformer, an innovative network that utilizes a hierarchical vision transformer as its backbone and integrates large language models (LLMs) to extract and embed clinical data in text-rich form. The model features a novel visual language attention module (VLAM) to combine visual and linguistic features, enabling language-aware visual encoding (LAVE). The Radformer was evaluated on a dataset of 2985 patients with head-and-neck cancer who underwent RT. Quantitative evaluations were performed utilizing metrics such as the Dice similarity coefficient (DSC), intersection over union (IOU), and 95th percentile Hausdorff distance (HD95). The Radformer demonstrated superior performance in segmenting RT target volumes compared to state-of-the-art models. On the head-and-neck cancer dataset, Radformer achieved a mean DSC of 0.76 ± 0.09 versus 0.66 ± 0.09, a mean IOU of 0.69 ± 0.08 versus 0.59 ± 0.07, and a mean HD95 of 7.82 ± 6.87 mm versus 14.28 ± 6.85 mm for gross tumor volume delineation, compared to the baseline 3D-UNETR. The Radformer model offers a clinically optimal means of RT target auto-delineation by integrating both imaging and clinical data through a visual language model. This approach improves the accuracy of RT target volume delineation, facilitating broader AI-assisted automation in RT treatment planning.

Assessment of hard tissue changes after horizontal guided bone regeneration with the aid of deep learning CBCT segmentation.

To investigate the performance of a deep learning (DL) model for segmenting cone-beam computed tomography (CBCT) scans taken before and after mandibular horizontal guided bone regeneration (GBR) to evaluate hard tissue changes. The proposed SegResNet-based DL model was trained on 70 CBCT scans. It was tested on 10 pairs of pre- and post-operative CBCT scans of patients who underwent mandibular horizontal GBR. DL segmentations were compared to semi-automated (SA) segmentations of the same scans. Augmented hard tissue segmentation performance was evaluated by spatially aligning pre- and post-operative CBCT scans and subtracting preoperative segmentations obtained by DL and SA segmentations from the respective postoperative segmentations. The performance of DL compared to SA segmentation was evaluated based on the Dice similarity coefficient (DSC), intersection over the union (IoU), Hausdorff distance (HD95), and volume comparison. The mean DSC and IoU between DL and SA segmentations were 0.96 ± 0.01 and 0.92 ± 0.02 in both pre- and post-operative CBCT scans. While HD95 values between DL and SA segmentations were 0.62mm ± 0.16mm and 0.77mm ± 0.31mm for pre- and post-operative CBCTs respectively. The DSC, IoU and HD95 averaged 0.85 ± 0.08; 0.78 ± 0.07 and 0.91 ± 0.92mm for augmented hard tissue models respectively. Volumes mandible- and augmented hard tissue segmentations did not differ significantly between the DL and SA methods. The SegResNet-based DL model accurately segmented CBCT scans acquired before and after mandibular horizontal GBR. However, the training database must be further increased to increase the model's robustness. Automated DL segmentation could aid treatment planning for GBR and subsequent implant placement procedures and in evaluating hard tissue changes.

SmartSeg: machine learning infrastructure software for accelerating medical image segmentation in patient-specific applications

ABSTRACT Patient-specific applications in biomechanics and orthopaedics call for segmentation of volumetric medical images. This task has historically been performed manually or semi-automatically, which entails significant effort by trained experts. Computer vision algorithms based on machine learning have emerged as an effective tool for accelerating medical image processing tasks through automation. The current work describes fully no-code software which supports development of custom machine learning models for automatic segmentation of volumetric medical images. This cloud-integrated desktop application allows researchers and biomedical engineers to define segmentation algorithms, configure annotated datasets, and train personalised machine learning algorithms with cloud compute in just a few clicks. A presented case study demonstrates an ideal workflow in the software using femoral cartilage segmentation from magnetic resonance imaging as a representative use case. Models developed in the software demonstrated mean dice similarity coefficients of up to 0.886 on a test cohort, which is competitive with previously reported methods developed with significantly larger data sets. The mean surface difference between digitised models reconstructed from ground truth and predicted segmentation was 0.33 mm. Results suggest the described software enables creation of accurate machine learning models with limited engineering effort.

Automatic Aortic Valve Extraction Using Deep Learning with Contrast-Enhanced Cardiac CT Images.

This study evaluates the use of deep learning techniques to automatically extract and delineate the aortic valve annulus region from contrast-enhanced cardiac CT images. Two approaches, namely, segmentation and object detection, were compared to determine their accuracy. A dataset of 32 contrast-enhanced cardiac CT scans was analyzed. The segmentation approach utilized the DeepLabv3+ model, while the object detection approach employed YOLOv2. The dataset was augmented through rotation and scaling, and five-fold cross-validation was applied. The accuracy of both methods was evaluated using the Dice similarity coefficient (DSC), and their performance in estimating the aortic valve annulus area was compared. The object detection approach achieved a mean DSC of 0.809, significantly outperforming the segmentation approach, which had a mean DSC of 0.711. Object detection also demonstrated higher precision and recall, with fewer false positives and negatives. The aortic valve annulus area estimation had a mean error of 2.55 mm. Object detection showed superior performance in identifying the aortic valve annulus region, suggesting its potential for clinical application in cardiac imaging. The results highlight the promise of deep learning in improving the accuracy and efficiency of preoperative planning for cardiovascular interventions.

Prior Knowledge-Guided U-Net for Automatic Clinical Target Volume Segmentation in Postmastectomy Radiation Therapy of Breast Cancer.

This study aimed to design and evaluate a prior-knowledge-guided U-Net (PK-UNet) for automatic clinical target volume (CTV) segmentation in postmastectomy radiation therapy for breast cancer. A total of 102 computed tomography (CT) scans from breast cancer patients who underwent postmastectomy were retrospectively collected. Of these, 80 scans were used for training with 5-fold cross-validation, and 22 scans for independent testing. The CTV included the chest wall, supraclavicular region, and axillary group III. The proposed PK-UNet method employs a 2-stage auto-segmentation process. Initially, the localization network categorizes CT slices based on the anatomic information of the CTV and generates prior knowledge labels. These outputs, along with the CT images, were fed into the final segmentation network. Quantitative evaluation was conducted using the mean Dice similarity coefficient (DSC), 95% Hausdorff distance, average surface distance, and surface DSC. A four-level objective scale evaluation was performed by 2 experienced radiation oncologists in a randomized double-blind manner. Quantitative evaluations revealed that PK-UNet significantly outperformed state-of-the-art segmentation methods (P < .01), with a mean DSC of 0.90 ± 0.02 and a 95% Hausdorff distance of 2.82 ± 1.29 mm. The mean average surface distance of PK-UNet was 0.91 ± 0.22 mm and the surface DSC was 0.84 ± 0.07, significantly surpassing the performance of AdwU-Net (P < .01) and showing comparable results to other models. Clinical evaluation confirmed the efficacy of PK-UNet, with 81.8% of the predicted contours being acceptable for clinical application. The advantages of the auto-segmentation capability of PK-UNet were most evident in the superior and inferior slices and slices with discontinuities at the junctions of different subregions. The average manual correction time was reduced to 1.02 min, compared with 18.20 min for manual contouring leading to a 94.4% reduction in working time. This study introduced the pioneering integration of prior medical knowledge into a deep learning framework for postmastectomy radiation therapy. This strategy addresses the challenges of CTV segmentation in postmastectomy radiation therapy and improves clinical workflow efficiency.

Early Detection of Macular Atrophy Automated Through 2D and 3D Unet Deep Learning.

Macular atrophy (MA) is an irreversible endpoint of age-related macular degeneration (AMD), which is the leading cause of blindness in the world. Early detection is therefore an unmet need. We have developed a novel automated method to identify MA in patients undergoing follow-up with optical coherence tomography (OCT) for AMD based on the combination of 2D and 3D Unet architecture. Our automated detection of MA relies on specific structural changes in OCT, including six established atrophy-associated lesions. Using 1241 volumetric OCTs from 125 eyes (89 patients), the performance of this combination Unet architecture is extremely encouraging, with a mean dice similarity coefficient score of 0.90 ± 0.14 and a mean F1 score of 0.89 ± 0.14. These promising results have indicated superiority when compared to human graders, with a mean similarity of 0.71 ± 0.27. We believe this deep learning-aided tool would be useful to monitor patients with AMD, enabling the early detection of MA and supporting clinical decisions.

Deep-learning assessment of hippocampal magnetic susceptibility in Alzheimer's disease.

Quantitative susceptibility mapping (QSM) is pivotal for analyzing neurodegenerative diseases. However, accurate hippocampal segmentation remains a challenge. This study introduces a method for extracting hippocampal magnetic susceptibility values using a convolutional neural network (CNN) model referred to as 3D residual UNET. The model was pre-trained on whole QSM images and hippocampal segmentations from 3D T1-weighted images of 297 patients with Alzheimer's disease and mild cognitive impairment. Fine-tuning was conducted through manually annotated hippocampal segmentations from the QSM images of 60 patients. The performance was assessed using the Dice similarity coefficient (DSC) and Pearson correlation coefficient. The developed model was applied to another 98 patients, 49 with AD and 49 with mild cognitive impairment (MCI), and the correlation between the hippocampal magnetic susceptibility and volume was evaluated. The mean DSC for the hippocampal segmentation model was 0.716 ± 0.045. The correlation coefficient between the magnetic susceptibility values derived from manual segmentation and the CNN model was 0.983. The Pearson correlation coefficient between magnetic susceptibility and hippocampal volume from the CNN model was -0.252 (p = 0.012) on the left side and -0.311 (p = 0.002) on the right. The 3D residual UNET model enhances hippocampal analysis precision using QSM, which is capable of accurately extracting magnetic susceptibility.

CT ventilation images produced by a 3D neural network show improvement over the Jacobian and HU DIR-based methods to predict quantized lung function.

Radiation-induced pneumonitis affects up to 33% of non-small cell lung cancer (NSCLC) patients, with fatal pneumonitis occurring in 2% of patients. Pneumonitis risk is related to the dose and volume of lung irradiated. Clinical radiotherapy plans assume lungs are functionally homogeneous, but evidence suggests that avoidance of high-functioning lung during radiotherapy can reduce the risk of radiation-induced pneumonitis. Radiotherapy avoidance structures can be constructed based on high-function regions indicated in a ventilation map, which can be produced from CT images. Existing methods of deriving such a CT ventilation image (CTVI) require the use of deformable image registration (DIR) of peak-inhale and -exhale CT images, which is susceptible to inaccuracy for small or low-intensity regions, and sensitive to image artefacts. To overcome these problems, we use a neural network to predict a ventilation map from breath-hold CT (BHCT). We used the nnU-Net pipeline to train five-fold cross-validated ensemble models to predict a ventilation map (CTVInnU-Net). The training data were comprised of registered BHCT and Galligas PET images from 20 patients. Three training sets were created to ensure performance was averaged over different test patients. For each set, images from two randomly selected test patients were set aside, and models were trained on the remaining images. The ground truth was established by quantizing the Galligas PET images, assigning each voxel a label of high-function(>70th percentile of intensity), medium-function (between 30th and 70th percentile), or low-function(<30th percentile). For comparison, we created a CTVI with a 2D U-Net (CTVInnU-Net-2D), and with the Jacobian (CTVIJac) and Hounsfield Units (CTVIHU) DIR-based methods which we quantized and labeled in the same way. The Dice similarity coefficient (DSC) and Hausdorff Distance 95th percentile (HD95) of each CTVI with the ground truth were measured separately for each lung function subregion. CTVInnU-Net had the highest similarity to the quantized Galligas PET with a mean (range) DSC over all three categories of lung function at 0.68 (0.56 to 0.82), compared with 0.64 (0.47 to 0.75) for CTVInnU-Net-2D, 0.60 (0.38 to 0.73) for CTVIJac, and 0.56 (0.30 to 0.75) for CTVIHU. CTVInnU-Net had the equal-lowest spatial distance to the quantized Galligas PET averaged over the three categories, with HD95 of 22mm (9 to 64mm), compared with 23mm (9 to 72mm) for CTVInnU-Net-2D, 22mm (12 to 63mm) for CTVIJac, and 26mm (12 to 58mm) for CTVIHU. Our 3D neural network produces a quantized CTVI with higher similarity to the ground truth than the 2D U-Net and DIR-based Jacobian and HU methods. As it produces a quantized CTVI directly, CTVInnU-Net avoids the need for thresholding to identify high-function lung regions. With faster evaluation and improved accuracy, neural networks show promise for the clinical implementation of functional lung avoidance.

Self-supervised learning on dual-sequence magnetic resonance imaging for automatic segmentation of nasopharyngeal carcinoma

Automating the segmentation of nasopharyngeal carcinoma (NPC) is crucial for therapeutic procedures but presents challenges given the hurdles in amassing extensively annotated datasets. Although previous studies have applied self-supervised learning to capitalize on unlabeled data to improve segmentation performance, these methods often overlooked the benefits of dual-sequence magnetic resonance imaging (MRI). In the present study, we incorporated self-supervised learning with a saliency transformation module using unlabeled dual-sequence MRI for accurate NPC segmentation. 44 labeled and 72 unlabeled patients were collected to develop and evaluate our network. Impressively, our network achieved a mean Dice similarity coefficient (DSC) of 0.77, which is consistent with a previous study that relied on a training set of 4,100 annotated cases. The results further revealed that our approach required minimal adjustments, primarily < 20% tweak in the DSC, to meet clinical standards. By enhancing the automatic segmentation of NPC, our method alleviates the annotation burden on oncologists, curbs subjectivity, and ensures reliable NPC delineation.

Imaging error reduction in radial cine-MRI with deep learning-based intra-frame motion compensation

Objective.Radial cine-MRI allows for sliding window reconstruction at nearly arbitrary frame rate, promising high-speed imaging for intra-fractional motion monitoring in magnetic resonance guided radiotherapy. However, motion within the reconstruction window may determine the location of the reconstructed target to deviate from the true real-time position (target positioning errors), particularly in cases of fast breathing or for anatomical structures affected by the heartbeat. In this work, we present a proof-of-concept study aiming to enhance radial cine-MR imaging by implementing deep-learning-based intra-frame motion compensation techniques.Approach.A novel network (TransSin-UNet) was proposed to continuously estimate the final-position image of the target, corresponding to end of the frame acquisition. Within the radial k-space reconstruction window, the spatial-temporal dependencies among the sinogram representation of the spokes were modeled by a transformer encoder subnetwork, followed by a UNet subnetwork operating in the spatial domain for pixel-level fine-tuning. By simulating motion-dependent radial sampling with (tiny) golden angles, we generated datasets from 25 4D digital anthropomorphic lung cancer phantoms. The network was then trained and extensively evaluated across datasets characterized by varying azimuthal radial profile increments.Main Results.The method required additional 4.8 ms per frame over the conventional approach involving direct image reconstruction with motion-corrupted spokes. TransSin-UNet outperformed architectures relying solely on transformer encoders or UNets across all the comparative evaluations, leading to a noticeable enhancement in image quality and target positioning accuracy. The normalized root mean-squared error decreased by 50% from the initial value of 0.188 on average, whereas the mean Dice similarity coefficient of the gross tumor volume increased from 85.1% to 96.2% in the investigated cases. Furthermore, the final-positions of anatomical structures undergoing substantial intra-frame deformations were precisely derived.Significance.The proposed approach enables an effective intra-frame motion compensation, offering an opportunity to reduce errors in radial cine-MR imaging for real-time motion management.

Using CT images to assist the segmentation of MR images via generalization: Segmentation of the renal parenchyma of renal carcinoma patients.

Developing deep learning models for segmenting medical images in multiple modalities with less data and annotation is an attractive and challenging task, which was previously discussed as being accomplished by complex external frameworks for bridging the gap between different modalities. Exploring the generalization ability of networks in medical images in different modalities could provide more simple and accessible methods, yet comprehensive testing could still be needed. To explore the feasibility and robustness of using computed tomography (CT) images to assist the segmentation of magnetic resonance (MR) images via the generalization, in the segmentation of renal parenchyma of renal cell carcinoma (RCC) patients. Nephrographic CT images and fat-suppressed T2-weighted (fs-T2W) images were retrospectively collected. The pure CT dataset included 116 CT images. Additionally, 240 MR images were randomly divided into subsets A and B. From subset A, three training datasets were constructed, each containing 40, 80, and 120 images, respectively. Similarly, three datasets were constructed from subset B. Subsequently, datasets with mixed modality were created by combining these pure MR datasets with the 116 CT images. The 3D-UNET models for segmenting the renal parenchyma in two steps were trained using these 13 datasets: segmenting kidneys and then the renal parenchyma. These models were evaluated in internal MR (n=120), CT (n=65) validation datasets, and an external validation dataset of CT (n=79), using the mean of the dice similarity coefficient (DSC). To demonstrate the robustness of generalization ability over different proportions of modalities, we compared the models trained with mixed modality in three different proportions and pure MR, using repeated measures analysis of variance (RM-ANOVA). We developed a renal parenchyma volume quantification tool by the trained models. The mean differences and Pearson correlation coefficients between the model segmentation volume and the ground truth segmentation volume were calculated for its evaluation. The mean DSCs of models trained with 116 data in CT in the validation of MR were 0.826, 0.842, and 0.953, respectively, for the predictions of kidney segmentation model on whole image, renal parenchymal segmentation model on kidneys with RCC and without RCC. For all models trained with mixed modality, the means of DSC were above 0.9, in all validations of CT and MR. According to the results of the comparison between models trained with mixed modality and pure MR, the means of DSC of the former were significantly greater or equal to the latter, at all three different proportions of modalities. The differences of volumes were all significantly lower than one-third of the volumetric quantification error of a previous method, and the Pearson correlation coefficients of volumes were all above 0.96 on kidneys with and without RCC of three validations. CT images could be used to assist the segmentation of MR images via the generalization, with or without the supervision of MR data. This ability showed acceptable robustness. A tool for accurately measuring renal parenchymal volume on CT and MR images was established.

Dual-attention transformer-based hybrid network for multi-modal medical image segmentation

Accurate medical image segmentation plays a vital role in clinical practice. Convolutional Neural Network and Transformer are mainstream architectures for this task. However, convolutional neural network lacks the ability of modeling global dependency while Transformer cannot extract local details. In this paper, we propose DATTNet, DualATTentionNetwork, an encoder-decoder deep learning model for medical image segmentation. DATTNet is exploited in hierarchical fashion with two novel components: (1) Dual Attention module is designed to model global dependency in spatial and channel dimensions. (2) Context Fusion Bridge is presented to remix the feature maps with multiple scales and construct their correlations. The experiments on ACDC, Synapse and Kvasir-SEG datasets are conducted to evaluate the performance of DATTNet. Our proposed model shows superior performance, effectiveness and robustness compared to SOTA methods, with mean Dice Similarity Coefficient scores of 92.2%, 84.5% and 89.1% on cardiac, abdominal organs and gastrointestinal poly segmentation tasks. The quantitative and qualitative results demonstrate that our proposed DATTNet attains favorable capability across different modalities (MRI, CT, and endoscopy) and can be generalized to various tasks. Therefore, it is envisaged as being potential for practicable clinical applications. The code has been released on https://github.com/MhZhang123/DATTNet/tree/main.

Dosimetric benefit and clinical feasibility of deep inspiration breath-hold and volumetric modulated arc therapy-based postmastectomy radiotherapy for left-sided breast cancer

To evaluate the dosimetric benefits and clinical feasibility of deep inspiratory breath-hold (DIBH) combined with volumetric modulated arc therapy (VMAT) in left-sided postmastectomy radiotherapy (PMRT). Eligible patients with left-sided breast cancer undergoing DIBH-based PMRT were prospectively included. Chest wall, supra/infraclavicular fossa, and/or internal mammary node irradiation (IMNI) were planned with a prescription dose of 43.5 Gy in 15 fractions. VMAT plans were designed on free breathing (FB)—and DIBH-CT to compare dosimetric parameters in heart, left anterior descending artery (LAD) and lung. Cone-beam computed tomography (CBCT) was performed before and after treatment to evaluate inter- and intra-fractional setup errors. Heart position and dose variations during treatment were estimated by fusing CBCT with DIBH-CT scans.Twenty patients were included with 10 receiving IMNI. In total, 193 pre-treatment and 39 pairs pre- and post-treatment CBCT scans were analyzed. The Dmean, Dmax, and V5−40 of the heart, LAD, and left lung were significantly lower in DIBH than FB (p < 0.05 for all), except for V5 of LAD (p = 0.167). The cardiopulmonary dosimetric benefits were maintained regardless of IMNI. The inter- and intra-fractional setup errors were < 0.3 cm; and the overall estimated PTV margins were < 1.0 cm. During treatment, the mean dice similarity coefficient of heart position and the mean ratio of heart Dmean between CBCT and DIBH-CT plans was 0.95 (0.88–1.00) and 100% (70.6–119.5%), respectively. DIBH-VMAT could effectively reduce the cardiopulmonary doses with acceptable reproducibility and stability in left-sided PMRT regardless of IMNI.

Clinical target volume (CTV) automatic delineation using deep learning network for cervical cancer radiotherapy: A study with external validation.

To explore the accuracy and feasibility of a proposed deep learning (DL) algorithm for clinical target volume (CTV) delineation in cervical cancer radiotherapy and evaluate whether it can perform well in external cervical cancer and endometrial cancer cases for generalization validation. A total of 332 patients were enrolled in this study. A state-of-the-art network called ResCANet, which added the cascademulti-scale convolution in the skip connections to eliminate semantic differences between different feature layers based on ResNet-UNet. The atrous spatial pyramid pooling in the deepest feature layer combined the semantic information of different receptive fields without losing information. A total of 236 cervical cancer cases were randomly grouped into 5-fold cross-training (n=189) and validation (n=47) cohorts. External validations were performed ina separate cohort of 54 cervical cancer and 42 endometrial cancer cases. The performances of the proposed network were evaluated by dice similarity coefficient (DSC), sensitivity (SEN), positive predictive value (PPV), 95% Hausdorff distance (95HD), and oncologist clinical score when comparing them with manual delineation in validation cohorts. In internal validation cohorts, the mean DSC, SEN, PPV, 95HD for ResCANet achieved 74.8%, 81.5%, 73.5%, and 10.5mm. In external independent validation cohorts, ResCANet achieved 73.4%, 72.9%, 75.3%, 12.5mm for cervical cancer cases and 77.1%, 81.1%, 75.5%, 10.3mm for endometrial cancer cases, respectively. The clinical assessment score showed that minor and no revisions (delineation time was shortened to within 30min) accountedforabout 85% of all cases in DL-aided automatic delineation. We demonstrated the problem of model generalizability for DL-based automatic delineation. The proposed network can improve the performance of automatic delineation for cervical cancer and shorten manual delineation time at no expense to quality. The network showed excellent clinical viability, which can also be even generalized for endometrial cancer with excellent performance.

Automatic detection and segmentation of lesions in 18 F-FDG PET/CT imaging of patients with Hodgkin lymphoma using 3D dense U-Net.

The accuracy of automatic tumor segmentation in PET/computed tomography (PET/CT) images is crucial for the effective treatment and monitoring of Hodgkin lymphoma. This study aims to address the challenges faced by certain segmentation algorithms in accurately differentiating lymphoma from normal organ uptakes due to PET image resolution and tumor heterogeneity. Variants of the encoder-decoder architectures are state-of-the-art models for image segmentation. Among these kinds of architectures, U-Net is one of the most famous and predominant for medical image segmentation. In this study, we propose a fully automatic approach for Hodgkin lymphoma segmentation that combines U-Net and DenseNet architectures to reduce network loss for very small lesions, which is trained using the Tversky loss function. The hypothesis is that the fusion of these two deep learning models can improve the accuracy and robustness of Hodgkin lymphoma segmentation. A dataset with 141 samples was used to train our proposed network. Also, to test and evaluate the proposed network, we allocated two separate datasets of 20 samples. We achieved 0.759 as the mean Dice similarity coefficient with a median value of 0.767, and interquartile range (0.647-0.837). A good agreement was observed between the ground truth of test images against the predicted volume with precision and recall scores of 0.798 and 0.763, respectively. This study demonstrates that the integration of U-Net and DenseNet architectures, along with the Tversky loss function, can significantly enhance the accuracy of Hodgkin lymphoma segmentation in PET/CT images compared to similar studies.