Performance of a Deep Learning Algorithm for Automated Segmentation and Quantification of Traumatic Pelvic Hematomas on CT.

Higher glucose on admission is associated with need for angioembolization in stable pelvic fracture

Extension–contraction transformation network for pancreas segmentation in abdominal CT scans

Segmentation of individual thigh muscles in MRI images is essential for monitoring neuromuscular diseases and quantifying relevant biomarkers such as fat fraction (FF). Deep learning approaches such as U-Net have demonstrated effectiveness in this field. However, the impact of reducing neural network complexity remains unexplored in the FF quantification in individual muscles. U-Net architectures with different complexities have been compared for the quantification of the fat fraction in each muscle group selected in the central part of the thigh region. The corresponding performance has been assessed in terms of Dice score (DSC) and FF quantification error. The database contained 1450 thigh images from 59 patients and 14 healthy subjects (age: 47 ± 17years, sex: 36F, 37M). Ten individual muscles were segmented in each image. The performance of each model was compared to nnU-Net, a complex architecture with 4.35 107 parameters, 12.8 Gigabytes of peak memory usage and 167h of training time. As expected, nnU-Net achieved the highest DSC (94.77 ± 0.13%). A simpler U-Net (5.81 105 parameters, 2.37 Gigabytes, 14h of training time) achieved a lower DSC but still above 90%. Surprisingly, both models achieved a comparable FF estimation. The poor correlation between observed DSC and FF indicates that less complex architectures, reducing GPU memory utilization and training time, can still accurately quantify FF.

Existing methods for automated breast ultrasound lesions detection and recognition tend to be based on multi-stage processing, such as preprocessing, filtering/denoising, segmentation and classification. The performance of these processes is dependent on the prior stages. To improve the current state of the art, we have proposed an end-to-end breast ultrasound lesions detection and recognition using a deep learning approach. We implemented a popular semantic segmentation framework, i.e. Fully Convolutional Network (FCN-AlexNet) for our experiment. To overcome data deficiency, we used a pre-trained model based on ImageNet and transfer learning. We validated our results on two datasets, which consist of a total of 113 malignant and 356 benign lesions. We assessed the performance of the model using the following split: 70p for training data, 10p for validation data, and 20p testing data. The results show that our proposed method performed better on benign lesions, with a Dice score of 0.6879, when compared to the malignant lesions with a Dice score of 0.5525. When considering the number of images with Dice score > 0.5, 79p of the benign lesions were successfully segmented and correctly recognised, while 65p of the malignant lesions were successfully segmented and correctly recognised. This paper provides the first end-to-end solution for breast ultrasound lesion recognition. The future challenges for the proposed approaches are to obtain additional datasets and customize the deep learning framework to improve the accuracy of this method.

597 Background: Automated breast tumor identification and segmentation in magnetic resonance imaging (MRI) is a difficult and crucial area of study in breast cancer research. Artificial intelligence (AI) models are increasingly being developed for automated localization of lesions in imaging studies to facilitate quantitative assessment of features for improved diagnostic, prognostic and predictive performance. Such models have had success in detecting breast cancers in mammography, ultrasound and CT, but few have achieved three-dimensional (3D) volumetric tumor segmentation from breast MRI. The purpose of this study was to apply two state-of-the-art AI – specifically, deep learning (DL) - algorithms to 3D MRI breast cancer data and identify the higher performing algorithm for precise segmentation of breast tumors. Methods: We evaluated pre-treatment, T1 post-gadolinium contrast enhanced breast MRI from 222 patients with known breast cancers (n = 262). Images were split into training (n = 142), validation (n = 36), and hold-out test (n = 44) datasets. Two DL algorithms, U-Net and VAE-UNet, were trained to classify tumors on the training dataset across 1000 epochs. The output for each is a precise localization and segmentation of each tumor at the pixel level from every MRI image. We evaluated the performance of each algorithm using 5-fold cross-validation and testing on the validation and test sets. We calculated a dice accuracy score for each model as the performance comparison metric. Results: The highest dice accuracy score achieved on the validation dataset by generic U-Net was 83.38%, with an average across 1000 epochs of 62.41%. The highest dice accuracy achieved on the validation dataset by VAE-UNet was 82.62%, with an average across epochs of 61.28%. On our test dataset, the highest dice accuracy score achieved by U-Net was 93.09%, with an average across epochs of 66.31%, and the highest accuracy score for VAE-UNet was 90.98%, with average across epochs of 50.47%. Although U-Net appeared to perform slightly better than VAE-Unet for most cases, there were distinct cases where VAE-UNet outperformed U-Net (dice score up to 59% better than U-Net). Subsequent analysis indicated that VAE-UNet preferentially outperforms U-Net for tumors with low sphericity (p = 0.001). Conclusions: Our results suggest that U-Net is well suited for segmenting breast tumors from breast MRI in most cases, but that VAE-UNet outperforms U-Net when the tumor shapes are less spherical. Our findings could inform the choice of DL algorithms in research and clinical endeavors that rely on accurate breast cancer tumor segmentation. In particular, these two tools could be configured to facilitate tumor assessment from breast MRI in the clinical setting for: breast cancer screening in high-risk patient populations, pre-surgical planning, and monitoring of treatment response.

We present an evaluation of different deep learning and machine learning approaches for tree health classification in the Black Forest, the Harz Mountains, and the Göttinger Forest on a unique, highly accurate tree-level dataset. The multispectral UAV data were collected from eight forest plots with diverse tree species, mostly conifers. As ground truth data (GTD), nearly 1500 tree polygons with related attribute information on the health status of the trees were used. This data were collected during extensive fieldwork using a mobile application and subsequent individual tree segmentation. Extensive preprocessing included normalization, NDVI calculations, data augmentation to deal with the underrepresented classes, and splitting the data into training, validation, and test sets. We conducted several experiments using a classical machine learning approach (random forests), as well as different convolutional neural networks (CNNs)—ResNet50, ResNet101, VGG16, and Inception-v3—on different datasets and classes to evaluate the potential of these algorithms for tree health classification. Our first experiment was a binary classifier of healthy and damaged trees, which did not consider the degree of damage or tree species. The best results of a 0.99 test accuracy and an F1 score of 0.99 were obtained with ResNet50 on four band composites using the red, green, blue, and infrared bands (RGBI images), while VGG16 had the worst performance, with an F1 score of only 0.78. In a second experiment, we also distinguished between coniferous and deciduous trees. The F1 scores ranged from 0.62 to 0.99, with the highest results obtained using ResNet101 on derived vegetation indices using the red edge band of the camera (NDVIre images). Finally, in a third experiment, we aimed at evaluating the degree of damage: healthy, slightly damaged, and medium or heavily damaged trees. Again, ResNet101 had the best performance, this time on RGBI images with a test accuracy of 0.98 and an average F1 score of 0.97. These results highlight the potential of CNNs to handle high-resolution multispectral UAV data for the early detection of damaged trees when good training data are available.

The accurate counting of laminated sheets, such as packing or printing sheets in industry, is extremely important because it greatly affects the economic cost. However, the different thicknesses, adhesion properties, and breakage points and the low contrast of sheets remain challenges to traditional counting methods based on image processing. This paper proposes a new stacked-sheet counting method with a deep learning approach using the U-Net architecture. A specific dataset according to the characteristics of stack side images is collected. The stripe of the center line of each sheet is used for semantic segmentation, and the complete side images of the slices are segmented via training with small image patches and testing with original large images. With this model, each pixel is classified by multi-layer convolution and deconvolution to determine whether it is the target object to be detected. After the model is trained, the test set is used to test the model, and a center region segmentation map based on the pixel points is obtained. By calculating the statistical median value of centerline points across different sections in these segmented images, the number of sheets can be obtained. Compared with traditional image algorithms in real product counting experiments, the proposed method can achieve better performance with higher accuracy and a lower error rate.

Accurately predicting the 28-day compressive strength of Portland cement clinkers is challenging due to their complex multiphase nature, especially when production processes and environmental conditions vary. To address this issue, the TransXRD model was introduced, which is a novel deep learning approach that combines the Transformer architecture with CNN design. 214 clinker samples were collected from 6 cement plants. The diffraction peak sequence of samples ranging from 10° to 55° was used as required input for the model. A combination of one-dimensional convolution is incorporated to extract local features, and self-attention mechanisms are leveraged to perform global feature integration. XRD data were directly utilized, bypassing additional processing to minimize signal loss. Compared to other DNN (deep neural network) models, the TransXRD model demonstrates promising performance, achieving a root mean square error (RMSE) of 1.15, a mean squared error (MSE) of 1.33, a mean absolute error (MAE) of 0.93, and a mean absolute percentage error (MAPE) of 1.76 on the test set. Notably, in the training set, 100% of the data exhibit prediction errors within 2 MPa, while within the test set, 90% of the prediction errors are within 2 MPa. The application potential of XRD data in predicting the compressive strength of cement clinker also offers novel insights into predicting the performance of other materials.

A dual-branch hybrid dilated CNN model for the AI-assisted segmentation of meningiomas in MR images

As of today, perioperative monitoring in the operating room is based on vital signs and clinical observations by the anesthesiologist. This, however, does not offer a complete monitoring of left ventricular function throughout the intervention. We hypothesize that functional monitoring of the heart can be performed automatically based on transesophageal echocardiographic (TEE) images. The main goal of this work is therefore to compute the non-linear deformation between subsequent images in a TEE sequence of the left ventricle and estimate basal longitudinal strain to assess regional myocardial function using a deep learning approach. An unsupervised approach based on a convolutional neural network (CNN) was implemented. The output of the CNN network is a dense vector field that describes the non-linear deformation required to maximize the similarity between two images.Recordings from 42 consecutive complete TEE exams from the Echocardiography Unit were anonymized and used for training. Recordings from 5 consecutive TEE exams performed during heart surgery, also anonymized, were used as test set. For the test set patients, the basal strain was manually annotated in EchoPac by an expert echocardiographer. For this work, 19 heart cycles were randomly selected from the test set and checked to ensure visibility of the points to be tracked. Overall when estimating strain there was a mean difference of 7.25%(±4.56%) in the strain values when compared to expert annotations.We have proven that point tracking is working as expected in most images, where the myocardium is well depicted. However, dropouts, noise generated by implants or air bubbles after surgery, confuse the tracker and drifting occurs.

The shape of pancreas in different patients is very different and the boundary is fuzzy, so the reliable automatic segmentation of pancreas is an important and difficult task. In this paper, we propose a multi-scale hybrid attention cascade network for challenging pancreas segmentation. First, the original CT images are coarsely segmented through the first-level network, then, the coarse segmentation results are clipped and sent to the second-level network for further training to obtain the fine segmentation results. The network adopts full convolutional network (FCN) integrating multi-scale mixed attention. The cut CT image eliminates the irrelevant background interference and reduces the input size of the secondary network, thus improving the segmentation accuracy. An extensive evaluation of 82 open data sets was performed by quadruple cross validation. The experimental results showed that the Dice coefficient was 84.62±4.20% compared with several advanced methods. In addition, this method has excellent performance in three measurement indexes of precision, Jaccard and recall.

Parkinson’s Disease (PD) is a neurological disorder that causes patients with Parkinson’s Disease (PPD) to have difficulty with body balancing. Thus, PPD rely on caretakers to fulfill their daily needs. Vision-based assistive systems could be useful for PPD to communicate with caretakers. The work presented here is a hybrid YOLOv8n-seg framework with DETR. In this framework, the traditional YOLOv8n-seg model’s head is replaced with DETR as the head for hand gesture (HG) segmentation for PPD. Since no public dataset exists for PPD gestures, a dataset of 4,583 raw hand gesture images was collected with a webcam under realistic home and clinical environments (such as poor light, cluttered background, and motion blur) and expanded via augmentation to 11,230 gestures. This dataset was divided into an 80% training set, a 15% validation set, and a 5% testing set with 9 classes (e.g., hungry, attention, call, toilet) to ensure robust evaluation. The baseline YOLOv8n-seg model and Transformer-based variant, DETR (DEtection TRansformer), were tested on the custom PPD hand gesture dataset. Compared to the baseline YOLOv8n-seg, the implemented hybrid model achieved superior performance across all evaluation metrics, with ~1% improvement in precision (99% vs. 98%), recall (97% vs. 96%), F1 score (98% vs. 97%), and dice score (98% vs. 97%), with almost the same mAP@50 (97% vs. 97% for all), while improving inference speed by +3.0% (55.1 FPS vs. 53.5 FPS). On the same custom dataset, the conventional U-Net achieved 88% precision, 92% recall, and a 0.9 dice score, whereas the proposed hybrid model reached 99% precision, 98% recall, and a 0.98 dice score. This confirms the superior performance of the hybrid model over the conventional U-Net architecture for HG segmentation. The Raspberry Pi 4B is used as an edge device for HGR of PPD. These enhancements demonstrate that the hybrid approach achieves both higher accuracy and faster real-time performance, which is useful for assistive systems deployment on the embedded edge device. To our knowledge, this is the first work combining YOLOv8n-seg with a DETR head for PPD hand gesture segmentation.

Automatic segmentation of abdomen organs using medical imaging has many potential applications in clinical workflows. Recently, the state-of-the-art performance for organ segmentation has been achieved by deep learning models, i.e., convolutional neural network (CNN). However, it is challenging to train the conventional CNN-based segmentation models that aware of the shape and topology of organs. In this work, we tackle this problem by introducing a novel end-to-end shape learning architecture -- organ point-network. It takes deep learning features as inputs and generates organ shape representations as points that located on organ surface. We later present a novel adversarial shape learning objective function to optimize the point-network to capture shape information better. We train the point-network together with a CNN-based segmentation model in a multi-task fashion so that the shared network parameters can benefit from both shape learning and segmentation tasks. We demonstrate our method with three challenging abdomen organs including liver, spleen, and pancreas. The point-network generates surface points with fine-grained details and it is found critical for improving organ segmentation. Consequently, the deep segmentation model is improved by the introduced shape learning as significantly better Dice scores are observed for spleen and pancreas segmentation.

Pancreas CT segmentation offers promise at understanding the structural manifestation of metabolic conditions. To date, the medical primary record of conditions that impact the pancreas is in the electronic health record (EHR) in terms of diagnostic phenotype data (e.g., ICD-10 codes). We posit that similar structural phenotypes could be revealed by studying subjects with similar medical outcomes. Segmentation is mainly driven by imaging data, but this direct approach may not consider differing canonical appearances with different underlying conditions (e.g., pancreatic atrophy versus pancreatic cysts). To this end, we exploit clinical features from EHR data to complement image features for enhancing the pancreas segmentation, especially in high-risk outcomes. Specifically, we propose, to the best of our knowledge, the first phenotype embedding model for pancreas segmentation by predicting representatives that share similar comorbidities. Such an embedding strategy can adaptively refine the segmentation outcome based on the discriminative contexts distilled from clinical features. Experiments with 2000 patients’ EHR data and 300 CT images with the healthy pancreas, type II diabetes, and pancreatitis subjects show that segmentation by predictive phenotyping significantly improves performance over state-of-the-arts (Dice score 0.775 to 0.791, \(p < 0.05\), Wilcoxon signed-rank test). The proposed method additionally achieves superior performance on two public testing datasets, BTCV MICCAI Challenge 2015 and TCIA pancreas CT. Our approach provides a promising direction of advancing segmentation with phenotype features while without requiring EHR data as input during testing.

(1) Background: The stethoscope is one of the main accessory tools in the diagnosis of temporomandibular joint disorders (TMD). However, the clinical auscultation of the masticatory system still lacks computer-aided support, which would decrease the time needed for each diagnosis. This can be achieved with digital signal processing and classification algorithms. The segmentation of acoustic signals is usually the first step in many sound processing methodologies. We postulate that it is possible to implement the automatic segmentation of the acoustic signals of the temporomandibular joint (TMJ), which can contribute to the development of advanced TMD classification algorithms. (2) Methods: In this paper, we compare two different methods for the segmentation of TMJ sounds which are used in diagnosis of the masticatory system. The first method is based solely on digital signal processing (DSP) and includes filtering and envelope calculation. The second method takes advantage of a deep learning approach established on a U-Net neural network, combined with long short-term memory (LSTM) architecture. (3) Results: Both developed methods were validated against our own TMJ sound database created from the signals recorded with an electronic stethoscope during a clinical diagnostic trail of TMJ. The Dice score of the DSP method was 0.86 and the sensitivity was 0.91; for the deep learning approach, Dice score was 0.85 and there was a sensitivity of 0.98. (4) Conclusions: The presented results indicate that with the use of signal processing and deep learning, it is possible to automatically segment the TMJ sounds into sections of diagnostic value. Such methods can provide representative data for the development of TMD classification algorithms.