Hum Brain Mapp. 2024 Feb 15;45(3):e26595. doi: 10.1002/hbm.26595.

ABSTRACT

Obesity is associated with negative effects on the brain. We exploit Artificial Intelligence (AI) tools to explore whether differences in clinical measurements following lifestyle interventions in overweight population could be reflected in brain morphology. In the DIRECT-PLUS clinical trial, participants with criterion for metabolic syndrome underwent an 18-month lifestyle intervention. Structural brain MRIs were acquired before and after the intervention. We utilized an ensemble learning framework to predict Body-Mass Index (BMI) scores, which correspond to adiposity-related clinical measurements from brain MRIs. We revealed that patient-specific reduction in BMI predictions was associated with actual weight loss and was significantly higher in active diet groups compared to a control group. Moreover, explainable AI (XAI) maps highlighted brain regions contributing to BMI predictions that were distinct from regions associated with age prediction. Our DIRECT-PLUS analysis results imply that predicted BMI and its reduction are unique neural biomarkers for obesity-related brain modifications and weight loss.

PMID:38375968 | DOI:10.1002/hbm.26595

Comput Methods Biomech Biomed Engin. 2024 Feb 20:1-18. doi: 10.1080/10255842.2023.2282951. Online ahead of print.

ABSTRACT

Drug discovery relies on the precise prognosis of drug-target interactions (DTI). Due to their ability to learn from raw data, deep learning (DL) methods have displayed outstanding performance over traditional approaches. However, challenges such as imbalanced data, noise, poor generalization, high cost, and time-consuming processes hinder progress in this field. To overcome the above challenges, we propose a DL-based model termed DrugSchizoNet for drug interaction (DI) prediction of Schizophrenia. Our model leverages drug-related data from the DrugBank and repoDB databases, employing three key preprocessing techniques. First, data cleaning eliminates duplicate or incomplete entries to ensure data integrity. Next, normalization is performed to enhance security and reduce costs associated with data acquisition. Finally, feature extraction is applied to improve the quality of input data. The three layers of the DrugSchizoNet model are the input, hidden and output layers. In the hidden layer, we employ dropout regularization to mitigate overfitting and improve generalization. The fully connected (FC) layer extracts relevant features, while the LSTM layer captures the sequential nature of DIs. In the output layer, our model provides confidence scores for potential DIs. To optimize the prediction accuracy, we utilize hyperparameter tuning through OB-MOA optimization. Experimental results demonstrate that DrugSchizoNet achieves a superior accuracy of 98.70%. The existing models, including CNN-RNN, DANN, CKA-MKL, DGAN, and CNN, across various evaluation metrics such as accuracy, recall, specificity, precision, F1 score, AUPR, and AUROC are compared with the proposed model. By effectively addressing the challenges of imbalanced data, noise, poor generalization, high cost and time-consuming processes, DrugSchizoNet offers a promising approach for accurate DTI prediction in Schizophrenia. Its superior performance demonstrates the potential of DL in advancing drug discovery and development processes.

PMID:38375638 | DOI:10.1080/10255842.2023.2282951

Comput Struct Biotechnol J. 2024 Feb 8;23:918-928. doi: 10.1016/j.csbj.2024.02.004. eCollection 2024 Dec.

ABSTRACT

The identification of photobleaching steps in single molecule fluorescence imaging is a well-established procedure for analysing the stoichiometries of molecular complexes. Nonetheless, the method is challenging with protein fluorophores because of the high levels of noise, rapid bleaching and highly variable signal intensities, all of which complicate methods based on statistical analyses of intensities to identify bleaching steps. It has recently been shown that deep learning by convolutional neural networks can yield an accurate analysis with a relatively short computational time. We describe here an improved use of such an approach that detects bleaching events even in the first time point of observation, and we have included this within an integrated software package incorporating fluorescence spot detection, colocalisation, tracking, FRET and photobleaching step analyses of single molecules or complexes. This package, known as FluoroTensor, is written in Python with a self-explanatory user interface.

PMID:38375530 | PMC:PMC10875188 | DOI:10.1016/j.csbj.2024.02.004

Front Neuroinform. 2024 Feb 5;18:1338886. doi: 10.3389/fninf.2024.1338886. eCollection 2024.

ABSTRACT

STUDY OBJECTIVES: We aimed to build a tool which facilitates manual labeling of sleep slow oscillations (SOs) and evaluate the performance of traditional sleep SO detection algorithms on such a manually labeled data set. We sought to develop improved methods for SO detection.

METHOD: SOs in polysomnographic recordings acquired during nap time from ten older adults were manually labeled using a custom built graphical user interface tool. Three automatic SO detection algorithms previously used in the literature were evaluated on this data set. Additional machine learning and deep learning algorithms were trained on the manually labeled data set.

RESULTS: Our custom built tool significantly decreased the time needed for manual labeling, allowing us to manually inspect 96,277 potential SO events. The three automatic SO detection algorithms showed relatively low accuracy (max. 61.08%), but results were qualitatively similar, with SO density and amplitude increasing with sleep depth. The machine learning and deep learning algorithms showed higher accuracy (best: 99.20%) while maintaining a low prediction time.

CONCLUSIONS: Accurate detection of SO events is important for investigating their role in memory consolidation. In this context, our tool and proposed methods can provide significant help in identifying these events.

PMID:38375447 | PMC:PMC10875054 | DOI:10.3389/fninf.2024.1338886

Heliyon. 2024 Feb 6;10(4):e25821. doi: 10.1016/j.heliyon.2024.e25821. eCollection 2024 Feb 29.

ABSTRACT

The global surge in energy demand, driven by technological advances and population growth, underscores the critical need for effective management of electricity supply and demand. In certain developing nations, a significant challenge arises because the energy demand of their population exceeds their capacity to generate, as is the case in Iraq. This study focuses on energy forecasting in Iraq, using a previously unstudied dataset from 2019 to 2021, sourced from the Iraqi Ministry of Electricity. The study employs a diverse set of advanced forecasting models, including Linear Regression, XGBoost, Random Forest, Long Short-Term Memory, Temporal Convolutional Networks, and Multi-Layer Perceptron, evaluating their performance across four distinct forecast horizons (24, 48, 72, and 168 hours ahead). Key findings reveal that Linear Regression is a consistent top performer in demand forecasting, while XGBoost excels in supply forecasting. Statistical analysis detects differences in models performances for both datasets, although no significant differences are found in pairwise comparisons for the supply dataset. This study emphasizes the importance of accurate energy forecasting for energy security, resource allocation, and policy-making in Iraq. It provides tools for decision-makers to address energy challenges, mitigate power shortages, and stimulate economic growth. It also encourages innovative forecasting methods, the use of external variables like weather and economic data, and region-specific models tailored to Iraq's energy landscape. The research contributes valuable insights into the dynamics of electricity supply and demand in Iraq and offers performance evaluations for better energy planning and management, ultimately promoting sustainable development and improving the quality of life for the Iraqi population.

PMID:38375305 | PMC:PMC10875426 | DOI:10.1016/j.heliyon.2024.e25821

Heliyon. 2024 Feb 8;10(4):e25960. doi: 10.1016/j.heliyon.2024.e25960. eCollection 2024 Feb 29.

ABSTRACT

Assigning medical codes for patients is essential for healthcare organizations, not only for billing purposes but also for maintaining accurate records of patients' medical histories and analyzing the outputs of certain procedures. Due to the abundance of disease codes, it can be laborious and time-consuming for medical specialists to manually assign these codes to each procedure. To address this problem, we discuss the automatic prediction of ICD-9 codes, the most popular and widely accepted system of medical coding. We introduce a two-stream deep learning framework specifically designed to analyze multi-modal data. This framework is applied to the extensive and publicly available MIMIC-III dataset, enabling us to leverage both numerical and text-based data for improved ICD-9 code prediction. Our system uses text representation models to understand the text-based medical records; the Gated Recurrent Unit (GRU) to model the numerical health records; and fuses these two streams to automatically predict the ICD-9 codes used in the intensive care unit. We discuss the preprocessing and classification methods and demonstrate that our proposed two-stream model outperforms other state-of-the-art studies in the literature.

PMID:38375292 | PMC:PMC10875443 | DOI:10.1016/j.heliyon.2024.e25960

Heliyon. 2024 Feb 3;10(4):e25844. doi: 10.1016/j.heliyon.2024.e25844. eCollection 2024 Feb 29.

ABSTRACT

In forensic medicine, estimating human skeletal remains' post-mortem interval (PMI) can be challenging. Following death, bones undergo a series of chemical and physical transformations due to their interactions with the surrounding environment. Post-mortem changes have been assessed using various methods, but estimating the PMI of skeletal remains could still be improved. We propose a new methodology with handheld hyperspectral imaging (HSI) system based on the first results from 104 human skeletal remains with PMIs ranging between 1 day and 2000 years. To differentiate between forensic and archaeological bone material, the Convolutional Neural Network analyzed 65.000 distinct diagnostic spectra: the classification accuracy was 0.58, 0.62, 0.73, 0.81, and 0.98 for PMIs of 0 week-2 weeks, 2 weeks-6 months, 6 months-1 year, 1 year-10 years, and >100 years, respectively. In conclusion, HSI can be used in forensic medicine to distinguish bone materials >100 years old from those <10 years old with an accuracy of 98%. The model has adequate predictive performance, and handheld HSI could serve as a novel approach to objectively and accurately determine the PMI of human skeletal remains.

PMID:38375262 | PMC:PMC10875450 | DOI:10.1016/j.heliyon.2024.e25844

iScience. 2024 Feb 1;27(3):109004. doi: 10.1016/j.isci.2024.109004. eCollection 2024 Mar 15.

ABSTRACT

Deep learning-based neuroimaging pipelines for acute stroke typically rely on image registration, which not only increases computation but also introduces a point of failure. In this paper, we propose a general-purpose contrastive self-supervised learning method that converts a convolutional deep neural network designed for registered images to work on a different input domain, i.e., with unregistered images. This is accomplished by using a self-supervised strategy that does not rely on labels, where the original model acts as a teacher and a new network as a student. Large vessel occlusion (LVO) detection experiments using computed tomographic angiography (CTA) data from 402 CTA patients show the student model achieving competitive LVO detection performance (area under the receiver operating characteristic curve [AUC] = 0.88 vs. AUC = 0.81) compared to the teacher model, even with unregistered images. The student model trained directly on unregistered images using standard supervised learning achieves an AUC = 0.63, highlighting the proposed method's efficacy in adapting models to different pipelines and domains.

PMID:38375230 | PMC:PMC10875112 | DOI:10.1016/j.isci.2024.109004

iScience. 2024 Feb 3;27(3):109105. doi: 10.1016/j.isci.2024.109105. eCollection 2024 Mar 15.

ABSTRACT

Wearable devices realize health monitoring, information transmission, etc. In this study, the human-friendliness, adaptability, reliability, and economy (HARE) principle for designing human energy harvesters is first proposed and then a biomechanical energy harvester (BMEH) is proposed to recover the knee negative energy to generate electricity. The proposed BMEH is mounted on the waist of the human body and connected to the ankles by ropes for driving. Double-rotor mechanism and half-wave rectification mechanism design effectively improves energy conversion efficiency with higher power output density for more stable power output. The experimental results demonstrate that the double-rotor mechanism increases the output power of the BMEH by 70% compared to the single magnet-rotor mechanism. And the output power density of BMEH reaches 0.07 W/kg at a speed of 7 km/h. Furthermore, the BMEH demonstrates the excitation mode detection accuracy of 99.8% based on the Gate Recurrent Unit deep learning model with optimal parameters.

PMID:38375224 | PMC:PMC10875156 | DOI:10.1016/j.isci.2024.109105

Front Endocrinol (Lausanne). 2024 Feb 2;15:1326009. doi: 10.3389/fendo.2024.1326009. eCollection 2024.

ABSTRACT

OBJECTIVE: The necessity of postmastectomy radiotherapy(PMRT) for patients with HR+/HER2 T1-2 N1M0 breast cancer remains controversial. We want to use explainable machine learning to learn the feature importance of the patients and identify the subgroup of the patients who may benefit from the PMRT. Additionally, develop tools to provide guidance to the doctors.

METHODS: In this study, we trained and validated 2 machine learning survival models: deep learning neural network and Cox proportional hazard model. The training dataset consisted of 35,347 patients with HR+/HER2- T1-2 N1M0 breast cancer who received mastectomies from the SEER database from 2013 to 2018. The performance of survival models were assessed using a concordance index (c-index).Then we did subgroup analysis to identify the subgroup who could benefit from PMRT. We also analyzed the global feature importance for the model and individual feature importance for individual survival prediction. Finally, we developed a Cloud-based recommendation system for PMRT to visualize the survival curve of each treatment plan and deployed it on the Internet.

RESULTS: A total of 35,347 patients were included in this study. We identified that radiotherapy improved the OS in patients with tumor size >14mm and age older than 54: 5-year OS rates of 91.9 versus 87.2% (radio vs. nonradio, P <0.001) and cohort with tumor size >14mm and grade worse than well-differentiated, 5-year OS rates of 90.8 versus 82.3% (radio vs. nonradio, P <0.001).The deep learning network performed more stably and accurately in predicting patients survival than the random survival forest and Cox proportional hazard model on the internal test dataset (C-index=0.776 vs 0.641) and in the external validation(C-index=0.769 vs 0.650).Besides, the deep learning model identified several key factors that significantly influence patient survival, including tumor size, examined regional nodes, age at 45-49 years old and positive regional nodes (PRN).

CONCLUSION: Patients with tumor size >14mm and age older than 54 and cohort with tumor size >14mm and grade worse than well-differentiated could benefit from the PMRT. The deep learning network performed more stably and accurately in predicting patients survival than Cox proportional hazard model on the internal test. Besides, tumor size, examined regional nodes, age at 45-49 years old and PRN are the most significant factors to the overall survival (OS).

PMID:38375194 | PMC:PMC10875455 | DOI:10.3389/fendo.2024.1326009

Front Artif Intell. 2024 Jan 11;6:1327355. doi: 10.3389/frai.2023.1327355. eCollection 2023.

ABSTRACT

Healthcare is a topic of significant concern within the academic and business sectors. The COVID-19 pandemic has had a considerable effect on the health of people worldwide. The rapid increase in cases adversely affects a nation's economy, public health, and residents' social and personal well-being. Improving the precision of COVID-19 infection forecasts can aid in making informed decisions regarding interventions, given the pandemic's harmful impact on numerous aspects of human life, such as health and the economy. This study aims to predict the number of confirmed COVID-19 cases in Saudi Arabia using Bayesian optimization (BOA) and deep learning (DL) methods. Two methods were assessed for their efficacy in predicting the occurrence of positive cases of COVID-19. The research employed data from confirmed COVID-19 cases in Saudi Arabia (SA), the United Kingdom (UK), and Tunisia (TU) from 2020 to 2021. The findings from the BOA model indicate that accurately predicting the number of COVID-19 positive cases is difficult due to the BOA projections needing to align with the assumptions. Thus, a DL approach was utilized to enhance the precision of COVID-19 positive case prediction in South Africa. The DQN model performed better than the BOA model when assessing RMSE and MAPE values. The model operates on a local server infrastructure, where the trained policy is transmitted solely to DQN. DQN formulated a reward function to amplify the efficiency of the DQN algorithm. By examining the rate of change and duration of sleep in the test data, this function can enhance the DQN model's training. Based on simulation findings, it can decrease the DQN work cycle by roughly 28% and diminish data overhead by more than 50% on average.

PMID:38375088 | PMC:PMC10875994 | DOI:10.3389/frai.2023.1327355

Phys Med Biol. 2024 Feb 19. doi: 10.1088/1361-6560/ad2a96. Online ahead of print.

ABSTRACT

This study aims to enhance medical image registration by addressing the limitations of existing approaches that rely on spatial transformations through U-Net, ConvNets, or Transformers. The objective is to develop a novel architecture that combines ConvNets, Graph Neural Networks (GNNs), and Capsule Networks to improve the accuracy and efficiency of medical image registration, which can also deal with the problem of rotating registration.&#xD;Approach: We propose an deep learning-based approach which can be utilized in both unsupervised and semi-supervised manners, named as HGCMorph. It leverages a hybrid framework that integrates ConvNets and GNNs to capture lower-level features, specifically short-range attention, while also utilizing Capsule Networks (CapsNets) to model abstract higher-level features, including entity properties such as position, size, orientation, deformation, and texture. This hybrid framework aims to provide a comprehensive representation of anatomical structures and their spatial relationships in medical images.&#xD;Main results: The results demonstrate the superiority of HGCMorph over existing state-of-the-art deep learning-based methods in both qualitative and quantitative evaluations. In unsupervised training process, our model outperforms the recent SOTA method TransMorph by achieving 7$\%$/38$\%$ increase on Dice Score Coefficient (DSC), and 2$\%$/7$\%$ improvement on Negative Jacobian Determinant (NJD) for OASIS and LPBA40 datasets, respectively. Furthermore, HGCMorph achieves improved registration accuracy in semi-supervised training process. In addition, when dealing with complex 3D rotations and secondary randomly deformations, our method still achieves the best performance. We also tested our methods on lung datasets, such as JSRT, Montgoermy and Shenzhen.&#xD;Significance: The significance lies in its innovative design to medical image registration. HGCMorph offers a novel framework that overcomes the limitations of existing methods by efficiently capturing both local and abstract features, leading to enhanced registration accuracy, discontinuity-preserving, and pose-learning abilities. The incorporation of Capsule Networks introduces valuable improvements, making the proposed method a valuable contribution to the field of medical image analysis. HGCMorph not only advances the SOTA methods but also has the potential to improve various medical applications that rely on accurate image registration.

PMID:38373349 | DOI:10.1088/1361-6560/ad2a96

Phys Med Biol. 2024 Feb 19. doi: 10.1088/1361-6560/ad2a97. Online ahead of print.

ABSTRACT

OBJECTIVE: Accurate delineation of organs-at-risk (OARs) is a critical step in radiotherapy. The deep learning generated segmentations usually need to be reviewed and corrected by oncologists manually, which is time-consuming and operator-dependent. Therefore, an automated quality assurance (QA) and adaptive optimization correction strategy was proposed to identify and optimize "incorrect" auto-segmentations.

APPROACH: A total of 586 CT images and labels from nine institutions were used. The OARs included the brainstem, parotid, and mandible. The deep learning generated contours were compared with the manual ground truth delineations. In this study, we proposed a novel Contour Quality Assurance and Adaptive Optimization (CQA-AO) strategy, which consists of the following three main components: 1) The contour QA module classified the deep learning generated contours as either accepted or unaccepted; 2) The unacceptable contour categories analysis module provided the potential error reasons (five unacceptable category) and locations (attention heatmaps); 3) The adaptive correction of unacceptable contours module integrate vision-language representations and utilize convex optimization algorithms to achieve adaptive correction of "incorrect" contours.

MAIN RESULTS: In the contour quality assurance tasks, the sensitivity (accuracy, precision) of CQA-AO strategy reached 0.940 (0.945, 0.948), 0.962 (0.937, 0.913), and 0.967 (0.962, 0.957) for brainstem, parotid and mandible, respectively. The unacceptable contour category analysis, the (F_I,〖Acc〗_I,F_micro,F_macro) of CQA-AO strategy reached (0.901,0.763,0.862,0.822), (0.855,0.737, 0.837, 0.784), and (0.907, 0.762, 0.858, 0.821) for brainstem, parotid and mandible, respectively. After adaptive optimization correction, the DSC values of brainstem, parotid and mandible have been improved by 9.4%, 25.9%, and 13.5%, and HD values decreased by 62%, 70.6%, and 81.6%, respectively.

SIGNIFICANCE: The proposed CQA-AO strategy, which combines quality assurance of contour and adaptive optimization correction for OARs contouring, demonstrated superior performance compare to conventional methods. This method can be implemented in the clinical contouring procedures and improve the efficiency of delineating and reviewing workflow.

PMID:38373347 | DOI:10.1088/1361-6560/ad2a97

Phys Med Biol. 2024 Feb 19. doi: 10.1088/1361-6560/ad2a94. Online ahead of print.

ABSTRACT

Computed Tomography (CT) has been widely used in industrial high-resolution non-destructive testing. However, it is difficult to obtain high-resolution images for large-scale objects due to their physical limitations. The objective is to develop an improved super-resolution technique that preserves small structures and details while efficiently capturing high-frequency information. The study proposes a new deep learning based method called spectrum learning (SPEAR) network for CT images super-resolution. This approach leverages both global information in the image domain and high-frequency information in the frequency domain. The SPEAR network is designed to reconstruct high-resolution images from low-resolution inputs by considering not only the main body of the images but also the small structures and other details. The symmetric property of the spectrum is exploited to reduce weight parameters in the frequency domain. Additionally, a spectrum loss is introduced to enforce the preservation of both high-frequency components and global information. The network is trained using pairs of low-resolution and high-resolution CT images, and it is fine-tuned using additional low-dose and normal-dose CT image pairs. The experimental results demonstrate that the proposed SPEAR network outperforms state-of-the-art networks in terms of image reconstruction quality. The approach successfully preserves high-frequency information and small structures, leading to better results compared to existing methods. The network's ability to generate high-resolution images from low-resolution inputs, even in cases of low-dose CT images, showcases its effectiveness in maintaining image quality. The proposed SPEAR network's ability to simultaneously capture global information and high-frequency details addresses the limitations of existing methods, resulting in more accurate and informative image reconstructions. This advancement can have substantial implications for various industries and medical diagnoses relying on accurate imaging.

PMID:38373346 | DOI:10.1088/1361-6560/ad2a94

Phys Med Biol. 2024 Feb 19. doi: 10.1088/1361-6560/ad2a95. Online ahead of print.

ABSTRACT

OBJECTIVE: Generally, due to a lack of explainability, radiomics based on deep learning has been perceived as a black-box solution for radiologists. Automatic generation of diagnostic reports is a semantic approach to enhance the explanation of deep learning radiomics (DLR).

APPROACH: In this paper, we propose a novel model called radiomics-reporting network (Radioport), which incorporates text attention. This model aims to improve the interpretability of deep learning radiomics in mammographic calcification diagnosis. Firstly, it employs convolutional neural networks (CNN) to extract visual features as radiomics for multi-category classification based on Breast Imaging Reporting and Data System (BI-RADS). Then, it builds a mapping between these visual features and textual features to generate diagnostic reports, incorporating an attention module for improved clarity.

MAIN RESULTS: To demonstrate the effectiveness of our proposed model, we conducted experiments on a breast calcification dataset comprising mammograms and diagnostic reports. The results demonstrate that our model can: (i) semantically enhance the interpretability of deep learning radiomics; and, (ii) improve the readability of generated medical reports.

SIGNIFICANCE: Our interpretable textual model can explicitly simulate the mammographic calcification diagnosis process.

PMID:38373345 | DOI:10.1088/1361-6560/ad2a95

Phys Med Biol. 2024 Feb 19. doi: 10.1088/1361-6560/ad2a99. Online ahead of print.

ABSTRACT

This study addresses a fundamental limitation of In-beam Positron Emission Tomography (IB-PET) in proton therapy: the lack of direct anatomical representation in the images it produces. We aim to overcome this shortcoming by pioneering the application of deep learning techniques to create synthetic control CT images (sCT) from combining IB-PET and planning CT scan data.&#xD;&#xD;Approach. We conducted simulations involving six patients who underwent irradiation with proton beams. Leveraging the architecture of a Visual Transformer (ViT) Neural Network (NN), we developed a model to generate sCT images of these patients using the planning CT scans and the inter-fractional simulated PET activity maps during irradiation. To evaluate the model's performance, a comparison was conducted between the sCT images produced by the ViT model and the authentic control CT images - serving as the benchmark.&#xD;&#xD;Main Results. The Structural Similarity Index (SSIM) was computed at a mean value across all patients of 0.91, while the Mean Absolute Error (MAE) measured 22 Hounsfield Units (HU). Root Mean Squared Error (RMSE) and Peak Signal-to-Noise Ratio (PSNR) values were 56 HU and 30 dB, respectively. The Dice Similarity Coefficient (DSC) exhibited a value of 0.98. These values are comparable to or exceed those found in the literature. More than 70% of the synthetic morphological changes were found to be geometrically compatible with the ones reported in the real control CT scan.&#xD;&#xD;Significance. Our study presents an innovative approach to surface the hidden anatomical information of IB-PET in proton therapy. Our ViT-based model successfully generates sCT images from inter-fractional PET data and planning CT scans. Our model's performance stands on par with existing models relying on input from Cone Beam CT (CBCT) or Magnetic Resonance Imaging (MRI), which contain more anatomical information than activity maps.

PMID:38373343 | DOI:10.1088/1361-6560/ad2a99

J Appl Clin Med Phys. 2024 Feb 19:e14298. doi: 10.1002/acm2.14298. Online ahead of print.

ABSTRACT

PURPOSE: Diagnosing Renal artery stenosis (RAS) presents challenges. This research aimed to develop a deep learning model for the computer-aided diagnosis of RAS, utilizing multimodal fusion technology based on ultrasound scanning images, spectral waveforms, and clinical information.

METHODS: A total of 1485 patients received renal artery ultrasonography from Peking Union Medical College Hospital were included and their color doppler sonography (CDS) images were classified according to anatomical site and left-right orientation. The RAS diagnosis was modeled as a process involving feature extraction and multimodal fusion. Three deep learning (DL) models (ResNeSt, ResNet, and XCiT) were trained on a multimodal dataset consisted of CDS images, spectrum waveform images, and individual basic information. Predicted performance of different models were compared with senior physician and evaluated on a test dataset (N = 117 patients) with renal artery angiography results.

RESULTS: Sample sizes of training and validation datasets were 3292 and 169 respectively. On test data (N = 676 samples), predicted accuracies of three DL models were more than 80% and the ResNeSt achieved the accuracy 83.49% ± 0.45%, precision 81.89% ± 3.00%, and recall 76.97% ± 3.7%. There was no significant difference between the accuracy of ResNeSt and ResNet (82.84% ± 1.52%), and the ResNeSt was higher than the XCiT (80.71% ± 2.23%, p < 0.05). Compared to the gold standard, renal artery angiography, the accuracy of ResNest model was 78.25% ± 1.62%, which was inferior to the senior physician (90.09%). Besides, compared to the multimodal fusion model, the performance of single-modal model on spectrum waveform images was relatively lower.

CONCLUSION: The DL multimodal fusion model shows promising results in assisting RAS diagnosis.

PMID:38373294 | DOI:10.1002/acm2.14298

J Appl Clin Med Phys. 2024 Feb 19:e14297. doi: 10.1002/acm2.14297. Online ahead of print.

ABSTRACT

PURPOSE: Deep learning-based auto-segmentation algorithms can improve clinical workflow by defining accurate regions of interest while reducing manual labor. Over the past decade, convolutional neural networks (CNNs) have become prominent in medical image segmentation applications. However, CNNs have limitations in learning long-range spatial dependencies due to the locality of the convolutional layers. Transformers were introduced to address this challenge. In transformers with self-attention mechanism, even the first layer of information processing makes connections between distant image locations. Our paper presents a novel framework that bridges these two unique techniques, CNNs and transformers, to segment the gross tumor volume (GTV) accurately and efficiently in computed tomography (CT) images of non-small cell-lung cancer (NSCLC) patients.

METHODS: Under this framework, input of multiple resolution images was used with multi-depth backbones to retain the benefits of high-resolution and low-resolution images in the deep learning architecture. Furthermore, a deformable transformer was utilized to learn the long-range dependency on the extracted features. To reduce computational complexity and to efficiently process multi-scale, multi-depth, high-resolution 3D images, this transformer pays attention to small key positions, which were identified by a self-attention mechanism. We evaluated the performance of the proposed framework on a NSCLC dataset which contains 563 training images and 113 test images. Our novel deep learning algorithm was benchmarked against five other similar deep learning models.

RESULTS: The experimental results indicate that our proposed framework outperforms other CNN-based, transformer-based, and hybrid methods in terms of Dice score (0.92) and Hausdorff Distance (1.33). Therefore, our proposed model could potentially improve the efficiency of auto-segmentation of early-stage NSCLC during the clinical workflow. This type of framework may potentially facilitate online adaptive radiotherapy, where an efficient auto-segmentation workflow is required.

CONCLUSIONS: Our deep learning framework, based on CNN and transformer, performs auto-segmentation efficiently and could potentially assist clinical radiotherapy workflow.

PMID:38373289 | DOI:10.1002/acm2.14297

J Appl Clin Med Phys. 2024 Feb 19:e14310. doi: 10.1002/acm2.14310. Online ahead of print.

ABSTRACT

PURPOSE: Radiation therapy (RT) of pediatric brain cancer is known to be associated with long-term neurocognitive deficits. Although target and organs-at-risk (OARs) are contoured as part of treatment planning, other structures linked to cognitive functions are often not included. This paper introduces a novel automatic segmentation tool specifically designed for the unique challenges posed by pediatric patients undergoing brain RT, as well as its seamless integration into the existing clinical workflow.

METHODS AND MATERIALS: Images of 47 pediatric brain cancer patients aged 1 to 20 years old and 33 two-year-old healthy infants were used to train a vision transformer, UNesT, for the segmentation of five brain OARs. The trained model was then incorporated to clinical workflow via DICOM connections between a treatment planning system (TPS) and a server hosting the trained model such that scans are sent from TPS to the server, automatically segmented, and sent back to TPS for treatment planning.

RESULTS: The proposed automatic segmentation framework achieved a median dice similarity coefficient of 0.928 (frontal white matter), 0.908 (corpus callosum), 0.933 (hippocampi), 0.819 (temporal lobes), and 0.960 (brainstem) with a mean ± SD run time of 1.8 ± 0.67 s over 20 test cases.

CONCLUSIONS: The pediatric brain segmentation tool showed promising performance on five OARs linked to neurocognitive functions and can easily be extended for additional structures. The proposed integration to the clinic enables easy access to the tool from clinical platforms and minimizes disruption to existing workflow while maximizing its benefits.

PMID:38373283 | DOI:10.1002/acm2.14310

Med Phys. 2024 Feb 19. doi: 10.1002/mp.16984. Online ahead of print.

ABSTRACT

BACKGROUND: Given the varying vulnerability of the rostral and caudal regions of the hippocampus to neuropathology in the Alzheimer's disease (AD) continuum, accurately assessing structural changes in these subregions is crucial for early AD detection. The development of reliable and robust automatic segmentation methods for hippocampal subregions (HS) is of utmost importance.

OBJECTIVE: Our aim is to propose and validate a HS segmentation model that is both training-free and highly generalizable. This method should exhibit comparable accuracy and efficiency to state-of-the-art techniques. The segmented HS can serve as a biomarker for studying the progression of AD.

METHODS: We utilized the functional magnetic resonance imaging of the Brain's Integrated Registration and Segmentation Tool (FIRST) to segment the entire hippocampus. By intersecting the segmentation results with the Brainnetome (BN) atlas, we obtained coarse segmentation of the four HS regions. This coarse segmentation was then employed as a shape prior term in the lattice Boltzmann (LB) model, as well as for initializing contours. Additionally, image gradients and local gray levels were integrated into the external force terms of the LB model to refine the coarse segmentation results. We assessed the segmentation accuracy of the model using the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset and evaluated the potential of the segmentation results as AD biomarkers on both the ADNI and Xuanwu datasets.

RESULTS: The median Dice similarity coefficients (DSC) for the left caudal, right caudal, left rostral, and right rostral hippocampus were 0.87, 0.88, 0.88, and 0.89, respectively. The proportion of segmentation results with a DSC exceeding 0.8 was 77%, 78%, 77%, and 94% for the respective regions. In terms of volume, the correlation coefficients between the segmentation results of the four HS regions and the gold standard were 0.95, 0.93, 0.96, and 0.96, respectively. Regarding asymmetry, the correlation coefficient between the segmentation result's right caudal minus left caudal and the corresponding gold standard was 0.91, while for right rostral minus left rostral, it was 0.93. Over time, we observed a decline in the volumes of the four HS regions and the total hippocampal volume of mild cognitive impairment (MCI) converters. Analysis of inter-group differences revealed that, except for the right rostral region in the ADNI dataset, the p-values for the four HS regions in the normal controls (NC), MCI, and AD groups from both datasets were all below 0.05. The right caudal hippocampal volume demonstrated correlation coefficients of 0.47 and 0.43 with the mini-mental state examination (MMSE) and Montreal cognitive assessment (MoCA), respectively. Similarly, the left rostral hippocampal volume showed correlation coefficients of 0.50 and 0.58 with MMSE and MoCA, respectively.

CONCLUSIONS: Our framework allows for direct application to different brain magnetic resonance (MR) datasets without the need for training. It eliminates the requirement for complex image preprocessing steps while achieving segmentation accuracy comparable to deep learning (DL) methods even with small sample sizes. Compared to traditional active contour models (ACM) and atlas-based methods, our approach exhibits significant speed advantages. The segmented HS regions hold promise as potential biomarkers for studying the progression of AD.

PMID:38373278 | DOI:10.1002/mp.16984