Average Dice Similarity Coefficient Research Articles

Deep Learning-Derived Quantitative Scores for Chronic Rhinosinusitis Assessment: Correlation With Quality of Life Outcomes.

Computed tomography (CT) plays a crucial role in assessing chronic rhinosinusitis, but lacks objective quantifiable indicators. This study aimed to use deep learning for automated sinus segmentation to generate distinct quantitative scores and explore their correlations with disease-specific quality of life. From July 2021 to August 2022, 445 CT data were collected from 2 medical centers. A deep learning model based on nnU-Net was trained for automatic sinus segmentation and internally validated using 300 cases. The remaining 145 cases were split into an external testing set (74 cases) and an independent testing set (71 cases). Two quantitative scores, the quantitative Lund-MacKay score and the quantitative opacification score (QOS), were derived from the segmentation results. The quantitative scores' efficacy was assessed by comparing them with the Lund-MacKay score (LMS), the 22-item Sinonasal Outcome Test score (SNOT-22), and other clinical variables through correlation analyses. Furthermore, the relationship between quantitative scores and postoperative quality of life improvement was explored using single-factor logistic regression. The segmentation model achieved average Dice similarity coefficients of 0.993, 0.978, 0.958, and 0.871 for the training, validation, external testing, and independent testing sets, respectively. Both quantitative scores significantly correlated with the LMS (rho = 0.87 and rho = 0.70, P < .001). Neither score correlated with the total SNOT-22 score, although the modified QOS showed significant correlations with the nasal and sleep subdomains (rho = 0.26 and rho = 0.27, P < .05). No significant association was found between quantitative score and postoperative improvement in quality of life. Deep learning enables the automated segmentation of sinuses on CT scans, producing quantitative scores of sinus opacification. These automatic quantitative scores may serve as tools for chronic rhinosinusitis assessment.

Multi-label segmentation of carpal bones in MRI using expansion transfer learning.

The purpose of this study was to develop a robust deep learning approach trained with a small in-vivo MRI dataset for multi-label segmentation of all eight carpal bones for therapy planning and wrist dynamic analysis.&#xD;Approach: A small dataset of 15 3.0-T MRI scans from five health subjects was employed within this study. The MRI data was variable with respect to the Field Of View (FOV), wide range of image intensity, and joint pose. A two-stage segmentation pipeline using modified 3D U-Net was proposed. In the first stage, a novel architecture, introduced as Expansion Transfer Learning (ETL), cascades the use of a focused Region Of Interest (ROI) cropped around ground truth for pretraining and a subsequent transfer by an expansion to the original FOV for a primary prediction. The bounding box around the ROI generated was utilized in the second stage for high-accuracy,&#xD;labeled segmentations of eight carpal bones. Different metrics including Dice Similarity Coefficient (DSC), Average Surface Distance (ASD) and Hausdorff Distance (HD) were used to evaluate performance between proposed and four state-of-the-art approaches.&#xD;Main results: With an average DSC of 87.8 %, an ASD of 0.46 mm, an average HD of 2.42mm in all datasets (96.1 %, 0.16 mm, 0.38mm in 12 datasets after exclusion criteria, respectively), the proposed approach showed an overall strongest performance than comparisons.&#xD;Significance: To our best knowledge, this is the first CNN-based multi-label segmentation approach for MRI human carpal bones. The ETL introduced in this work improved the ability to localize a small ROI in a large FOV. Overall, the interplay of a two-stage approach and ETL culminated in convincingly accurate segmentation scores despite a very small amount of image data.

Semi-Supervised Learning Allows for Improved Segmentation With Reduced Annotations of Brain Metastases Using Multicenter MRI Data.

Deep learning-based segmentation of brain metastases relies on large amounts of fully annotated data by domain experts. Semi-supervised learning offers potential efficient methods to improve model performance without excessive annotation burden. This work tests the viability of semi-supervision for brain metastases segmentation. Retrospective. There were 156, 65, 324, and 200 labeled scans from four institutions and 519 unlabeled scans from a single institution. All subjects included in the study had diagnosed with brain metastases. 1.5 T and 3 T, 2D and 3D T1-weighted pre- and post-contrast, and fluid-attenuated inversion recovery (FLAIR). Three semi-supervision methods (mean teacher, cross-pseudo supervision, and interpolation consistency training) were adapted with the U-Net architecture. The three semi-supervised methods were compared to their respective supervised baseline on the full and half-sized training. Evaluation was performed on a multinational test set from four different institutions using 5-fold cross-validation. Method performance was evaluated by the following: the number of false-positive predictions, the number of true positive predictions, the 95th Hausdorff distance, and the Dice similarity coefficient (DSC). Significance was tested using a paired samples t test for a single fold, and across all folds within a given cohort. Semi-supervision outperformed the supervised baseline for all sites with the best-performing semi-supervised method achieved an on average DSC improvement of 6.3% ± 1.6%, 8.2% ± 3.8%, 8.6% ± 2.6%, and 15.4% ± 1.4%, when trained on half the dataset and 3.6% ± 0.7%, 2.0% ± 1.5%, 1.8% ± 5.7%, and 4.7% ± 1.7%, compared to the supervised baseline on four test cohorts. In addition, in three of four datasets, the semi-supervised training produced equal or better results than the supervised models trained on twice the labeled data. Semi-supervised learning allows for improved segmentation performance over the supervised baseline, and the improvement was particularly notable for independent external test sets when trained on small amounts of labeled data. Artificial intelligence requires extensive datasets with large amounts of annotated data from medical experts which can be difficult to acquire due to the large workload. To compensate for this, it is possible to utilize large amounts of un-annotated clinical data in addition to annotated data. However, this method has not been widely tested for the most common intracranial brain tumor, brain metastases. This study shows that this approach allows for data efficient deep learning models across multiple institutions with different clinical protocols and scanners. 3 TECHNICAL EFFICACY: Stage 2.

Deep Learning-Based Segmentation of Cervical Posterior Longitudinal Ligament Ossification in Computed Tomography Images and Assessment of Spinal Cord Compression: A Two-Center Study.

This study aims to develop a fully automated, computed tomography (CT)-based deep learning (DL) model to segment ossified lesions of the posterior longitudinal ligament and to measure the thickness of the ossified material and calculate the cervical spinal cord compression factor. A total of 307 patients were enrolled, with 260 patients from Shanghai Changzheng Hospital, and 47 patients from the Traditional Chinese Medicine Hospital of Southwest Medical University. CT images were used to manually segment the Ossification of the posterior longitudinal ligament (OPLL) by 4 experienced radiologists. The DL model employing a 3-dimensional U-Net framework was developed to segment the OPLLs. The system also measures the thickness of the ossified material at its thickest point and the diameter of the spinal canal at the corresponding level. Segmentation performance was evaluated using the Dice Similarity Coefficient, Average Surface Distance, and intraclass correlation coefficient (ICC) between ground truth and segmentation volumes. Concordance between the radiologists' and the DL system's measurements of the ossified material thickness, residual spinal canal diameter at maximum compression, and cervical spinal cord compression coefficient was assessed in a randomly selected subset of 30 cases from the training set using ICCs and Bland-Altman plots. The DL system demonstrated average Dice Similarity Coefficient of 0.81, 0.75, and 0.71 for the training, internal validation, and external test sets, respectively. The mean Average Surface Distance was 1.30 for the training set, 2.35 for the internal validation set, and 2.63 for the external test set. The ICC values of 0.958 for the thickness of the ossified material and 0.974 for the residual canal diameter measurement. The proposed DL model effectively detects and separates ossification foci in OPLL on CT images. It exhibits comparable performance to radiologists in quantifying spinal cord compression metrics.

Using the coefficient of determination to identify injury regions after stroke in pre-clinical FDG-PET images

Background:In the analysis of brain fluorodeoxyglucose positron emission tomography (FDG-PET) images, intensity normalization is a necessary step to reduce inter-subject variability. However, the choice of the most appropriate normalization method in stroke studies remains unclear, as demonstrated by inconsistent findings in the literature. Materials and methods:Here, we propose a regression- and single-subject-based model for analyzing FDG-PET images without intensity normalization. Two independent data sets were collected before and after middle cerebral artery occlusion (MCAO), with one comprising 120 rats and the other 96 rats. After data preprocessing, voxel intensities in the same region and hemisphere were paired before and after the MCAO scan. A linear regression model was applied to the paired data, and the coefficient of determination R2 was calculated to measure the linearity. The R2 values between the ipsilateral and contralateral hemispheres were compared, and significant regions were defined as those with reduced linearity. Our method was compared with voxel-wise analysis under different intensity normalization methods and validated using the triphenyl tetrazolium chloride (TTC) staining data. Results:The significant regions identified by the proposed method showed a large degree of overlap with the infarcted regions identified by TTC data, as measured by the Dice similarity coefficient (DSC). The average DSC of the proposed method was 59.7%, whereas the DSCs of the existing approaches ranged from 41.4%∼51.3%. Additional validation using receiver operating characteristic (ROC) demonstrated that the area under the curve (AUC) of the average ROC curves reached 0.84 using the proposed method, whereas existing methods achieved AUCs ranging from 0.77∼0.79. The identified regions were consistent across the two independent data sets, and some findings were corroborated by other publications. Conclusions:The proposed model presents a novel quantitative approach for identifying injury regions post-stroke using FDG-PET images. The calculation does not require intensity normalization and can be applied to individual subjects. The method yields more sensitive results compared to existing identification methods.

A device-dependent auto-segmentation method based on combined generalized and single-device datasets.

Although generalized-dataset-based auto-segmentation models that consider various computed tomography (CT) scanners have shown great clinical potential, their application to medical images from unseen scanners remains challenging because of device-dependent image features. This study aims to investigate the performance of a device-dependent auto-segmentation model based on a combined dataset of a generalized dataset and single CT scanner dataset. We constructed two training datasets for 21 chest and abdominal organs. The generalized dataset comprised 1203 publicly available multi-scanner data. The device-dependent dataset comprised 1253 data, including the 1203 multi-CT scanner data and 50 single CT scanner data. Using these datasets, the generalized-dataset-based model (GDSM) and the device-dependent-dataset-based model (DDSM) were trained. The models were trained using nnU-Net and tested on ten data samples from a single CT scanner. The evaluation metrics included the Dice similarity coefficient (DSC), the Hausdorff distance (HD), and the average symmetric surface distance (ASSD), which were used to assess the overall performance of the models. In addition, DSCdiff, HDratio, and ASSDratio, which are variations of the three existing metrics, were used to compare the performance of the models across different organs. For the average DSC, the GDSM and DDSM had values of 0.9251 and 0.9323, respectively. For the average HD, the GDSM and DDSM had values of 10.66 and 9.139mm, respectively; for the average ASSD, the GDSM and DDSM had values of 0.8318 and 0.6656mm, respectively. Compared with the GDSM, the DDSM showed consistent performance improvements of 0.78%, 14%, and 20% for the DSC, HD, and ASSD metrics, respectively. In addition, compared with the GDSM, the DDSM had better DSCdiff values in 14 of 21 tested organs, better HDratio values in 13 of 21 tested organs, and better ASSDratio values in 14 of 21 tested organs. The three averages of the variant metrics were all better for the DDSM than for the GDSM. The results suggest that combining the generalized dataset with a single scanner dataset resulted in an overall improvement in model performance for that device image.

Comparison of Vendor-Pretrained and Custom-Trained Deep Learning Segmentation Models for Head-and-Neck, Breast, and Prostate Cancers.

We assessed the influence of local patients and clinical characteristics on the performance of commercial deep learning (DL) segmentation models for head-and-neck (HN), breast, and prostate cancers. Clinical computed tomography (CT) scans and clinically approved contours of 210 patients (53 HN, 49 left breast, 55 right breast, and 53 prostate cancer) were used to train and validate segmentation models integrated within a vendor-supplied DL training toolkit and to assess the performance of both vendor-pretrained and custom-trained models. Four custom models (HN, left breast, right breast, and prostate) were trained and validated with 30 (training)/5 (validation) HN, 34/5 left breast, 39/5 right breast, and 30/5 prostate patients to auto-segment a total of 24 organs at risk (OARs). Subsequently, both vendor-pretrained and custom-trained models were tested on the remaining patients from each group. Auto-segmented contours were evaluated by comparing them with clinically approved contours via the Dice similarity coefficient (DSC) and mean surface distance (MSD). The performance of the left and right breast models was assessed jointly according to ipsilateral/contralateral locations. The average DSCs for all structures in vendor-pretrained and custom-trained models were as follows: 0.81 ± 0.12 and 0.86 ± 0.11 in HN; 0.67 ± 0.16 and 0.80 ± 0.11 in the breast; and 0.87 ± 0.09 and 0.92 ± 0.06 in the prostate. The corresponding average MSDs were 0.81 ± 0.76 mm and 0.76 ± 0.56 mm (HN), 4.85 ± 2.44 mm and 2.42 ± 1.49 mm (breast), and 2.17 ± 1.39 mm and 1.21 ± 1.00 mm (prostate). Notably, custom-trained models showed significant improvements over vendor-pretrained models for 14 of 24 OARs, reflecting the influence of data/contouring variations in segmentation performance. These findings underscore the substantial impact of institutional preferences and clinical practices on the implementation of vendor-pretrained models. We also found that a relatively small amount of institutional data was sufficient to train customized segmentation models with sufficient accuracy.

Encoding matching criteria for cross-domain deformable image registration.

Most existing deep learning-based registration methods are trained on single-type images to address same-domain tasks, resulting in performance degradation when applied to new scenarios. Retraining a model for new scenarios requires extra time and data. Therefore, efficient and accurate solutions for cross-domain deformable registration are indemand. We argue that the tailor-made matching criteria in traditional registration methods is one of the main reason they are applicable in different domains. Motivated by this, we devise a registration-oriented encoder to model the matching criteria of image features and structural features, which is beneficial to boost registration accuracy andadaptability. Specifically, a general feature encoder (Encoder-G) is proposed to capture comprehensive medical image features, while a structural feature encoder (Encoder-S) is designed to encode the structural self-similarity into the global representation. Moreover, by updating Encoder-S using one-shot learning, our method can effectively adapt to different domains. The efficacy of our method is evaluated using MRI images from three different domains, including brain images (training/testing: 870/90 pairs), abdomen images (training/testing: 1406/90 pairs), and cardiac images (training/testing: 64770/870 pairs). The comparison methods include traditional method (SyN) and cutting-edge deep networks. The evaluation metrics contain dice similarity coefficient (DSC) and average symmetric surface distance (ASSD). In the single-domain task, our method attains an average DSC of 68.9%/65.2%/72.8%, and ASSD of 9.75/3.82/1.30 mm on abdomen/cardiac/brain images, outperforming the second-best comparison methods by large margins. In the cross-domain task, without one-shot optimization, our method outperforms other deep networks in five out of six cross-domain scenarios and even surpasses symmetric image normalization method (SyN) in two scenarios. By conducting the one-shot optimization, our method successfully surpasses SyN in all six cross-domainscenarios. Our method yields favorable results in the single-domain task while ensuring improved generalization and adaptation performance in the cross-domain task, showing its feasibility for the challenging cross-domain registration applications. The code is publicly available at https://github.com/JuliusWang-7/EncoderReg.

Robust Automated Mouse Micro-CT Segmentation Using Swin UNEt TRansformers.

Image-guided mouse irradiation is essential to understand interventions involving radiation prior to human studies. Our objective is to employ Swin UNEt TRansformers (Swin UNETR) to segment native micro-CT and contrast-enhanced micro-CT scans and benchmark the results against 3D no-new-Net (nnU-Net). Swin UNETR reformulates mouse organ segmentation as a sequence-to-sequence prediction task using a hierarchical Swin Transformer encoder to extract features at five resolution levels, and it connects to a Fully Convolutional Neural Network (FCNN)-based decoder via skip connections. The models were trained and evaluated on open datasets, with data separation based on individual mice. Further evaluation on an external mouse dataset acquired on a different micro-CT with lower kVp and higher imaging noise was also employed to assess model robustness and generalizability. The results indicate that Swin UNETR consistently outperforms nnU-Net and AIMOS in terms of the average dice similarity coefficient (DSC) and the Hausdorff distance (HD95p), except in two mice for intestine contouring. This superior performance is especially evident in the external dataset, confirming the model's robustness to variations in imaging conditions, including noise and quality, and thereby positioning Swin UNETR as a highly generalizable and efficient tool for automated contouring in pre-clinical workflows.

Automatic Segmentation of the Left Ventricle in Apical Four-Chamber View on Transesophageal Echocardiography Based on UNeXt Deep Neural Network.

Background/Objectives: The automatic left ventricle segmentation in transesophageal echocardiography (TEE) is of significant importance. In this paper, we constructed a large-scale TEE apical four-chamber view (A4CV) image dataset and proposed an automatic left ventricular segmentation method for the TEE A4CV based on the UNeXt deep neural network. Methods: UNeXt, a variant of U-Net integrating a multilayer perceptron, was employed for left ventricle segmentation in the TEE A4CV because it could yield promising segmentation performance while reducing both the number of network parameters and computational complexity. We also compared the proposed method with U-Net, TransUNet, and Attention U-Net models. Standard TEE A4CV videos were collected from 60 patients undergoing cardiac surgery, from the onset of anesthesia to the conclusion of the procedure. After preprocessing, a dataset comprising 3000 TEE images and their corresponding labels was generated. The dataset was randomly divided into training, validation, and test sets in an 8:1:1 ratio on the patient level. The training and validation sets were used to train the UNeXt, U-Net, TransUNet, and Attention U-Net models for left ventricular segmentation. The dice similarity coefficient (DSC) and Intersection over Union (IoU) were used to evaluate the segmentation performance of each model, and the Kruskal-Wallis test was employed to analyze the significance of DSC differences. Results: On the test set, the UNeXt model achieved an average DSC of 88.60%, outperforming U-Net (87.76%), TransUNet (85.75%, p < 0.05), and Attention U-Net (79.98%; p < 0.05). Additionally, the UNeXt model had a smaller number of parameters (1.47 million) and floating point operations (2.28 giga) as well as a shorter average inference time per image (141.73 ms), compared to U-Net (185.12 ms), TransUNet (209.08 ms), and Attention U-Net (201.13 ms). The average IoU of UNeXt (77.60%) was also higher than that of U-Net (76.61%), TransUNet (77.35%), and Attention U-Net (68.86%). Conclusions: This study pioneered the construction of a large-scale TEE A4CV dataset and the application of UNeXt to left ventricle segmentation in the TEE A4CV. The proposed method may be used for automatic segmentation of the left ventricle in the TEE A4CV.

A Semi-Supervised Multi-Region Segmentation Framework of Bladder Wall and Tumor with Wall-Enhanced Self-Supervised Pre-Training.

Bladder cancer is a prevalent and highly recurrent malignancy within the urinary tract. The accurate segmentation of the bladder wall and tumor in magnetic resonance imaging (MRI) is a crucial step in distinguishing between non-muscle-invasive and muscle-invasive types of bladder cancer, which plays a pivotal role in guiding clinical treatment decisions and influencing postoperative quality of life. The performance of data-driven methods is highly dependent on the quality of the annotations and datasets, however the amount of high-quality annotated data is very limited given the difficulty of professional radiologists to distinguish the mixed regions between the bladder wall and the tumor. The performance of the data-driven approach is highly dependent on the quality of the annotation and datasets, Therefore, in order to alleviate these problems and take full advantage of the potential of limited annotated and unlabeled data, we designed a semi-supervised multi-region framework for bladder wall and tumor segmentation. Our framework incorporates wall-enhanced self-supervised pre-training, designed to enhance discrimination of the bladder wall, and a semi-supervised segmentation network that utilizes both limited high-quality annotated data and unlabeled data. Contrast consistency and reconstruction observation losses are introduced to constrain the model to enhance the bladder walls, and adaptive learning rate and post-processing techniques are implemented to further improve segmentation performance. Extensive experimental validation demonstrated that our proposed method achieves promising results in the segmentation of both the bladder wall and the tumor. The average Dice Similarity Coefficients (DSCs) of the proposed method for the bladder wall and tumor were 0.8351 and 0.9175, respectively. Visualization results indicated that our method can effectively reduce excessive segmentation artifacts outside the bladder, and improve the clinical significance of the segmentation results.

An improved multi-scale feature extraction network for medical image segmentation.

The use of U-Net and its variations has led to significant advancements in medical image segmentation. However, the encoder-decoder structures of these models often lose spatial information during downsampling. Skip connections can help address this issue; however, they may also introduce excessive irrelevant background information. Additionally, medical images display significant scale variations and complex tissue structures, making it challenging for existing models to accurately separate tissues from the background. To address these issues, we developed the Res2Net-ConvFormer-Dilation-UNet (Res2-CD-UNet), a multi-scale feature extraction network for medical image segmentation. This study presents a novel U-shaped segmentation network that employs Res2Net as the backbone and incorporates a convolution-style transformer in the encoding stage to enhance global attention. Additionally, a novel channel feature fusion block (CFFB) has been introduced in the skip connection stage to minimize the effects of background noise. The proposed model was evaluated using publicly available datasets, Synapse and Seg.A.2023. Using the Synapse dataset, the average dice similarity coefficient (DSC) reached 83.92%, which was 1.96% higher than the suboptimal model, and the average Hausdorff distance (HD) was 14.51 mm. Among the eight organs evaluated, optimal results were achieved for four organs. Similarly, using the Seg.A.2023 dataset, the proposed model also achieved the best results with an average DSC of 93.27%. The results of this study indicate that the proposed model can more accurately segment regions of interest and better extract multi-scale features in medical images than existing deep-learning algorithms.

Deep Learning-Based Automatic Segmentation Combined with Radiomics to Predict Post-TACE Liver Failure in HCC Patients.

To develop and validate a deep learning-based automatic segmentation model and combine with radiomics to predict post-TACE liver failure (PTLF) in hepatocellular carcinoma (HCC) patients. This was a retrospective study enrolled 210 TACE-trated HCC patients. Automatic segmentation model based on nnU-Net neural network was developed to segment medical images and assessed by the Dice similarity coefficient (DSC). The screened clinical and radiomics variables were separately used to developed clinical and radiomics predictive model, and were combined through multivariate logistic regression analysis to develop a combined predictive model. The area under the curve (AUC), calibration curve, and decision curve analysis (DCA) were applied to compare the performance of the three predictive models. The automatic segmentation model showed satisfactory segmentation performance with an average DSC of 83.05% for tumor segmentation and 92.72% for non-tumoral liver parenchyma segmentation. The international normalized ratio (INR) and albumin (ALB) was identified as clinically independent predictors for PTLF and used to develop clinical predictive model. Ten most valuable radiomics features, including 8 from non-tumoral liver parenchyma and 2 from tumor, were selected to develop radiomics predictive model and to calculate Radscore. The combined predictive model achieved the best and significantly improved predictive performance (AUC: 0.878) compared to the clinical predictive model (AUC: 0.785) and the radiomics predictive model (AUC: 0.815). This reliable combined predictive model can accurately predict PTLF in HCC patients, which can be a valuable reference for doctors in making suitable treatment plan.

Automatic jawbone structure segmentation on dental CBCT images via deep learning.

This study developed and evaluated a two-stage deep learning-based system for automatic segmentation of mandibular cortical bone, mandibular cancellous bone, maxillary cortical bone and maxillary cancellous bone on cone beam computed tomography (CBCT) images. A dataset containing 155 CBCT scans acquired with different parameters was obtained. A two-stage deep learning-based system was developed for automatically segmenting jawbone structures. The Dice similarity coefficient (DSC) and average symmetric surface distance (ASSD) were used to assess the segmentation performance of the system by comparing the automatic segmentation results with the ground truth. The impact of dental and quality abnormalities on segmentation performance was analysed, and a comparison of automatic segmentation (AS) with manually refined segmentation (MRS) was reported. The system achieved promising segmentation performance, with average DSC values of 93.69%, 96.83%, 86.14% and 95.57% and average ASSD values of 0.13mm, 0.16mm, 0.29mm and 0.41mm for the mandibular cortical bone, mandibular cancellous bone, maxillary cortical bone and maxillary cancellous bone, respectively. Quality abnormalities had a negative impact on segmentation performance. The performance metrics (DSCs > 98.8% and ASSDs < 0.1mm) indicated high overlap between the AS and MRS. The proposed system offers an accurate and time-efficient method for segmenting jawbone structures on CBCT images. Automatically segmenting jawbone structures is essential in most digital dental workflows. The proposed system has considerable potential for application in digital clinical workflows to assist dentists in making more accurate diagnoses and developing patient-specific treatment plans.

Bilateral symmetry-based augmentation method for improved tooth segmentation in panoramic X-rays

Panoramic X-rays are crucial in dental radiology, providing detailed images that are essential for diagnosing and planning treatment for various oral conditions. The advent of automated methods that learn from annotated data promises to significantly aid clinical experts in making accurate diagnoses. However, these methods often require large amounts of annotated data, making the generation of high-quality annotations for panoramic X-rays both challenging and time-consuming. This paper introduces a novel bilateral symmetry-based augmentation method specifically designed to enhance tooth segmentation in panoramic X-rays. By exploiting the inherent bilateral symmetry of these images, our proposed method systematically generates augmented data, leading to substantial improvements in the performance of tooth segmentation models. By increasing the training data size fourfold, our approach proportionately reduces the effort required to manually annotate extensive datasets. These findings highlight the potential of leveraging the symmetrical properties of medical images to enhance model performance and accuracy in dental radiology. The effectiveness of the proposed method is evaluated on three widely adopted deep learning models: U-Net, SE U-Net, and TransUNet. Significant improvements in segmentation accuracy are observed with the proposed augmentation method across all models. For example, the average Dice Similarity Coefficient (DSC) increases by over 8%, reaching 76.7% for TransUNet. Further, comparisons with existing augmentation methods, including rigid transform-based and elastic grid-based techniques, show that the proposed method consistently outperforms them with additional improvements up to 5% in terms of average DSC, with the exact improvement varying depending on the model and training dataset size. We have made the data augmentation codes and tools developed based on our method available at https://github.com/wathoresanket/bilateralsymmetrybasedaugmentation.

MBF-Net: Multi-scale boundary-aware aggregation for bi-directional information exchange and feature reshaping for medical image segmentation

As a critical component of computer-aided diagnosis systems, medical image segmentation plays a vital role in assisting clinicians in making rapid and accurate decisions and formulating treatment plans. Nevertheless, precise medical image segmentation still presents a number of challenges, including insufficient feature extraction capabilities in the presence of limited sample sizes, blurred segmentation boundaries, and information loss between the encoder and decoder. In order to address these issues, we propose a Multi-Scale Boundary-Aware Aggregation Network with Bidirectional Information Exchange and Feature Refinement (MBF-Net) for medical image segmentation. Initially, we design a Multi-Scale Boundary-Aware Aggregation Encoder (MBAE) that aggregates features from different scales and pixel levels within the input images, capturing fine-grained boundary information in deep features and establishing comprehensive global and local multi-scale contextual dependencies. This design significantly enhances the model's understanding of the overall image structure and its ability to discern subtle differences between lesions and background. Subsequently, a Multi-Scale Bidirectional Information Transmission (MBIT) module is introduced, which integrates bidirectional information flow between low-level and high-level features, enabling multi-scale features to flow bidirectionally across different layers. The MBIT module effectively preserves crucial boundary details during cross-layer information transmission, thereby bridging the semantic gap between the encoder and decoder, and thereby improving the clarity of the segmentation boundaries. Finally, we develop a Feature Refinement and Aggregation Fusion (FRAF) module, designed to integrate feature information from various semantic levels, which alleviates discrepancies between features at varying scales, thus enhancing the segmentation accuracy of the network. The generalisation and effectiveness of MBF-Net are validated through comprehensive experiments on a range of tasks, including nuclear segmentation, breast cancer segmentation, polyp segmentation and skin lesion segmentation. Both subjective and objective evaluations demonstrate that MBF-Net significantly outperforms current state-of-the-art methods, achieving average Dice Similarity Coefficient (DSC) and Intersection over Union (IoU) scores of 86.34 % and 78.37 %, respectively. The superior performance of MBF-Net in terms of segmentation accuracy and quality is demonstrated across five public datasets.

Detection of fetal congenital heart defects on three-vessel view ultrasound videos

Background:Detecting congenital heart defects (CHDs) is challenging due to the difficulty of identifying subtle abnormalities in fetal heart structures. Objectives:To develop a deep learning-based method for segmenting vessels in the three-vessel view (3VV) to characterise the vessels by size and spatial relationships to detect abnormal fetal hearts. Methods:We present a deep learning-based method that takes as input a fetal heart ultrasound (US) video of the three vessels view (3VV) and an anchor frame, which contains the segmentation of the pulmonary artery (PA), aorta (Ao), and superior vena cava (SVC) in the 3VV. The method automatically segments the anatomical structures subsequent to the anchor frame and classifies the US video as normal or abnormal. The method consists of two phases. The first phase combines three residual networks (ResNets) extended with a self-attention block and a refinement module. The second phase extends a ResNet with two CoordConv layers integrating spatial coordinates. We assess segmentation performance using the intersection over union (IoU) and dice similarity coefficient (DSC) metrics and classification of US videos using sensitivity and specificity. We also investigate the tolerance to failure of the method by introducing mislabelled anchor frames. The dataset used in this study consists of 150 US videos of the 3VV; 50 videos were used for training, and 100 videos (50 normal videos, 50 abnormal videos) for testing. Results:In terms of anatomical structure segmentation accuracy, the method achieves an average IoU of 89.5% (99.5% for PA, 85.0% for Ao, and 84.1% for SVC), and an average DSC of 0.950% (0.946% for PA, 0.969% for Ao, and 0.934% for SVC). Detection of abnormal videos achieved a sensitivity of 0.99 and specificity of 1.0. The tolerance to failure analysis shows a decrease in the sensitivity of 0.023 and 0.015 for normal and abnormal case videos, respectively. Conclusions:The initial evaluation of our approach to fetal CHDs on 3VV ultrasound videos is promising but requires further refinement and evaluation on a larger dataset to assess clinical utility. The approach is designed to be translatable to low-resource settings where fetal echocardiography experts are unavailable due to the simple acquisition protocol.

Radiomics-guided generative adversarial network for automatic primary target volume segmentation for nasopharyngeal carcinoma using computed tomography images.

Automatic primary gross tumor volume (GTVp) segmentation for nasopharyngeal carcinoma (NPC) is a quite challenging task because of the existence of similar visual characteristics between tumors and their surroundings, especially on computed tomography (CT) images with severe low contrast resolution. Therefore, most recently proposed methods based on radiomics or deep learning (DL) is difficult to achieve good results on CT datasets. A peritumoral radiomics-guided generative adversarial network (PRG-GAN) was proposed to address this challenge. A total of 157 NPC patients with CT images was collected and divided into training, validation, and testing cohorts of 108, 9, and 30 patients, respectively. The proposed model was based on a standard GAN consisting of a generator network and a discriminator network. Morphological dilation on the initial segmentation results from GAN was first conducted to delineate annular peritumoral region, in which radiomics features were extracted as priori guide knowledge. Then, radiomics features were fused with semantic features by the discriminator's fully connected layer to achieve the voxel-level classification and segmentation. The dice similarity coefficient (DSC), 95% Hausdorff distance (HD95), and average symmetric surface distance (ASSD) were used to evaluate the segmentation performance using a paired samples t-test with Bonferroni correction and Cohen's d (d) effect sizes. A two-sided p-value of less than 0.05 was considered statistically significant. The model-generated predictions had a high overlap ratio with the ground truth. The average DSC, HD95, and ASSD were significantly improved from 0.80±0.12, 4.65±4.71mm, and 1.35±1.15mm of GAN to 0.85±0.18 (p=0.001, d=0.71), 4.15±7.56mm (p=0.002, d=0.67), and 1.11±1.65mm (p<0.001, d=0.46) of PRG-GAN, respectively. Integrating radiomics features into GAN is promising to solve unclear border limitations and increase the delineation accuracy of GTVp for patients with NPC.

Automated Vertebral Bone Quality Determination from T1-Weighted Lumbar Spine MRI Data Using a Hybrid Convolutional Neural Network–Transformer Neural Network

Vertebral bone quality (VBQ) is a promising new method that can improve screening for osteoporosis. The drawback of the current method is that it requires manual determination of the regions of interest (ROIs) of vertebrae and cerebrospinal fluid (CSF) by a radiologist. In this work, an automatic method for determining the VBQ is proposed, in which the ROIs are obtained using a trained neural network model. A large, publicly available dataset of sagittal lumbar spine MRI images with ground truth segmentations was used to train a BRAU-Net++ hybrid CNN–transformer neural network. The performance of the trained model was evaluated using the dice similarity coefficient (DSC), accuracy, precision, recall and intersection-over-union (IoU) metrics. The trained model performed similarly to state-of-the-art lumbar spine segmentation models, with an average DSC value of 0.914 ± 0.007 for the vertebrae and 0.902 for the spinal canal. Four different methods of VBQ determination with automatic segmentation are presented and compared with one-way ANOVA. These methods use different algorithms for CSF extraction from the segmentation of the spinal canal using T1- and T2-weighted image data and applying erosion to the vertebral ROI to avoid a sharp change in SI at the edge of the vertebral body.

An automated pheochromocytoma and paraganglioma lesion segmentation AI-model at whole-body 68Ga- DOTATATE PET/CT.

Somatostatin receptor (SSR) targeting radiotracer 68Ga-DOTATATE is used for Positron Emission Tomography (PET)/Computed Tomography (CT) imaging to assess patients with Pheochromocytoma and paraganglioma (PPGL), rare types of Neuroendocrine tumor (NET) which can metastasize thereby becoming difficult to quantify. The goal of this study is to develop an artificial intelligence (AI) model for automated lesion segmentation on whole-body 3D DOTATATE-PET/CT and to automate the tumor burden calculation. 132 68Ga-DOTATATE PET/CT scans from 38 patients with metastatic and inoperable PPGL, were split into 70, and 62 scans, from 20, and 18 patients for training, and test sets, respectively. The training set was further divided into patient-stratified 5 folds for cross-validation. 3D-full resolution nnUNet configuration was trained with 5-fold cross-validation. The model's detection performance was evaluated at both scan and lesion levels for the PPGL test set and two other clinical cohorts with NET (n = 9) and olfactory neuroblastoma (ONB, n = 5). Additionally, quantitative statistical analysis of PET parameters including SUVmax, total lesion uptake (TLU), and total tumor volume (TTV), was conducted. The nnUNet AI model achieved an average 5-fold validation dice similarity coefficient of 0.84 at the scan level. The model achieved dice similarity coefficients (DSC) of 0.88, 0.6, and 0.67 at the scan level, the sensitivity of 86%, 61.13%, and 61.64%, and a positive predictive value of 89%, 74%, and 86.54% at the lesion level for the PPGL test, NET and ONB cohorts, respectively. For PPGL cohorts, smaller lesions with low uptake were missed by the AI model (p < 0.001). Anatomical region-based failure analysis showed most of the false negative and false positive lesions within the liver for all the cohorts, mainly due to the high physiologic liver background activity and image noise on 68Ga- DOTATATE PET scans. The developed deep learning-based AI model showed reliable performance for automated segmentation of metastatic PPGL lesions on whole-body 68Ga-DOTATATE-PET/CT images, which may be beneficial for tumor burden estimation for objective evaluation during therapy follow-up. https://www. gov/study/NCT03206060 , https://www. gov/study/NCT04086485 , https://www. gov/study/NCT05012098 .