Nat Commun. 2024 Feb 29;15(1):1853. doi: 10.1038/s41467-024-46089-y.

ABSTRACT

Many machine learning applications in bioinformatics currently rely on matching gene identities when analyzing input gene signatures and fail to take advantage of preexisting knowledge about gene functions. To further enable comparative analysis of OMICS datasets, including target deconvolution and mechanism of action studies, we develop an approach that represents gene signatures projected onto their biological functions, instead of their identities, similar to how the word2vec technique works in natural language processing. We develop the Functional Representation of Gene Signatures (FRoGS) approach by training a deep learning model and demonstrate that its application to the Broad Institute's L1000 datasets results in more effective compound-target predictions than models based on gene identities alone. By integrating additional pharmacological activity data sources, FRoGS significantly increases the number of high-quality compound-target predictions relative to existing approaches, many of which are supported by in silico and/or experimental evidence. These results underscore the general utility of FRoGS in machine learning-based bioinformatics applications. Prediction networks pre-equipped with the knowledge of gene functions may help uncover new relationships among gene signatures acquired by large-scale OMICs studies on compounds, cell types, disease models, and patient cohorts.

PMID:38424040 | DOI:10.1038/s41467-024-46089-y

J Nucl Med. 2024 Feb 29:jnumed.123.267048. doi: 10.2967/jnumed.123.267048. Online ahead of print.

ABSTRACT

Automatic detection and characterization of cancer are important clinical needs to optimize early treatment. We developed a deep, semisupervised transfer learning approach for fully automated, whole-body tumor segmentation and prognosis on PET/CT. Methods: This retrospective study consisted of 611 18F-FDG PET/CT scans of patients with lung cancer, melanoma, lymphoma, head and neck cancer, and breast cancer and 408 prostate-specific membrane antigen (PSMA) PET/CT scans of patients with prostate cancer. The approach had a nnU-net backbone and learned the segmentation task on 18F-FDG and PSMA PET/CT images using limited annotations and radiomics analysis. True-positive rate and Dice similarity coefficient were assessed to evaluate segmentation performance. Prognostic models were developed using imaging measures extracted from predicted segmentations to perform risk stratification of prostate cancer based on follow-up prostate-specific antigen levels, survival estimation of head and neck cancer by the Kaplan-Meier method and Cox regression analysis, and pathologic complete response prediction of breast cancer after neoadjuvant chemotherapy. Overall accuracy and area under the receiver-operating-characteristic (AUC) curve were assessed. Results: Our approach yielded median true-positive rates of 0.75, 0.85, 0.87, and 0.75 and median Dice similarity coefficients of 0.81, 0.76, 0.83, and 0.73 for patients with lung cancer, melanoma, lymphoma, and prostate cancer, respectively, on the tumor segmentation task. The risk model for prostate cancer yielded an overall accuracy of 0.83 and an AUC of 0.86. Patients classified as low- to intermediate- and high-risk had mean follow-up prostate-specific antigen levels of 18.61 and 727.46 ng/mL, respectively (P < 0.05). The risk score for head and neck cancer was significantly associated with overall survival by univariable and multivariable Cox regression analyses (P < 0.05). Predictive models for breast cancer predicted pathologic complete response using only pretherapy imaging measures and both pre- and posttherapy measures with accuracies of 0.72 and 0.84 and AUCs of 0.72 and 0.76, respectively. Conclusion: The proposed approach demonstrated accurate tumor segmentation and prognosis in patients across 6 cancer types on 18F-FDG and PSMA PET/CT scans.

PMID:38423786 | DOI:10.2967/jnumed.123.267048

AJNR Am J Neuroradiol. 2024 Feb 29. doi: 10.3174/ajnr.A8173. Online ahead of print.

ABSTRACT

BACKGROUND AND PURPOSE: Spontaneous intracranial hypotension is an increasingly recognized condition. Spontaneous intracranial hypotension is caused by a CSF leak, which is commonly related to a CSF-venous fistula. In patients with spontaneous intracranial hypotension, multiple intracranial abnormalities can be observed on brain MR imaging, including dural enhancement, "brain sag," and pituitary engorgement. This study seeks to create a deep learning model for the accurate diagnosis of CSF-venous fistulas via brain MR imaging.

MATERIALS AND METHODS: A review of patients with clinically suspected spontaneous intracranial hypotension who underwent digital subtraction myelogram imaging preceded by brain MR imaging was performed. The patients were categorized as having a definite CSF-venous fistula, no fistula, or indeterminate findings on a digital subtraction myelogram. The data set was split into 5 folds at the patient level and stratified by label. A 5-fold cross-validation was then used to evaluate the reliability of the model. The predictive value of the model to identify patients with a CSF leak was assessed by using the area under the receiver operating characteristic curve for each validation fold.

RESULTS: There were 129 patients were included in this study. The median age was 54 years, and 66 (51.2%) had a CSF-venous fistula. In discriminating between positive and negative cases for CSF-venous fistulas, the classifier demonstrated an average area under the receiver operating characteristic curve of 0.8668 with a standard deviation of 0.0254 across the folds.

CONCLUSIONS: This study developed a deep learning model that can predict the presence of a spinal CSF-venous fistula based on brain MR imaging in patients with suspected spontaneous intracranial hypotension. However, further model refinement and external validation are necessary before clinical adoption. This research highlights the substantial potential of deep learning in diagnosing CSF-venous fistulas by using brain MR imaging.

PMID:38423747 | DOI:10.3174/ajnr.A8173

J Biomed Inform. 2024 Feb 27:104616. doi: 10.1016/j.jbi.2024.104616. Online ahead of print.

ABSTRACT

OBJECTIVE: This study aims to comprehensively review the use of graph neural networks (GNNs) for clinical risk prediction based on electronic health records (EHRs). The primary goal is to provide an overview of the state-of-the-art of this subject, highlighting ongoing research efforts and identifying existing challenges in developing effective GNNs for improved prediction of clinical risks.

METHODS: A search was conducted in the Scopus, PubMed, ACM Digital Library, and Embase databases to identify relevant English-language papers that used GNNs for clinical risk prediction based on EHR data. The study includes original research papers published between January 2009 and May 2023.

RESULTS: Following the initial screening process, 50 articles were included in the data collection. A significant increase in publications from 2020 was observed, with most selected papers focusing on diagnosis prediction (n = 36). The study revealed that graph attention networks (GAT) (n = 19) was the most prevalent architecture, and MIMIC-III (n = 23) was the most common data resource.

CONCLUSION: GNNs are relevant tools for predicting clinical risk by accounting for the relational aspects among medical events and entities and managing large volumes of EHR data. Future studies in this area may address challenges such as EHR data heterogeneity, multimodality, and model interpretability, aiming to develop more holistic GNN models that can produce more accurate predictions, be effectively implemented in clinical settings, and ultimately improve patient care.

PMID:38423267 | DOI:10.1016/j.jbi.2024.104616

Photodiagnosis Photodyn Ther. 2024 Feb 27:104030. doi: 10.1016/j.pdpdt.2024.104030. Online ahead of print.

ABSTRACT

BACKGROUND: Portwine stain (PWS) birthmarks are congenital vascular malformations. The quantification of PWS area is an important step in lesion classification and treatment evaluation.

AIMS: The aim of this study was to evaluate the combination of 3D scan with deep learning for automated PWS area quantization.

MATERIALS AND METHODS: PWS color was measured using a portable spectrophotometer. PWS patches (29.26 - 45.82 cm2) of different color and shape were generated for 2D and 3D PWS model. 3D images were acquired by a handheld 3D scanner to create texture maps. For semantic segmentation, an improved DeepLabV3+ network was developed for PWS lesion extraction from texture mapping of 3D images. In order to achieve accurate extraction of lesion regions, the convolutional block attention module (CBAM) and DENSE were introduced and the network was trained under Ranger optimizer. The performance of different backbone networks for PWS lesion extraction were also compared.

RESULTS: IDeepLabV3+ (Xception) showed the best results in PWS lesion extraction and area quantification. Its mean Intersection over Union (MIou) was 0.9797, Mean Pixel Accuracy (MPA) 0.9908, Accuracy 0.9989, Recall 0.9886 and F1-score 0.9897, respectively. In PWS area quantization, the mean value of the area error rate of this scheme was 2.61 ± 2.34.

CONCLUSIONS: The new 3D method developed in this study was able to achieve accurate quantification of PWS lesion area and has potentials for clinical applications.

PMID:38423233 | DOI:10.1016/j.pdpdt.2024.104030

Ergonomics. 2024 Feb 29:1-15. doi: 10.1080/00140139.2024.2306315. Online ahead of print.

ABSTRACT

Preventing work-related musculoskeletal disorders (WMSDs) is crucial in reducing their impact on individuals and society. However, the existing mainstream 2D image-based approach is insufficient in capturing the complex 3D movements and postures involved in many occupational tasks. To address this, an improved deep learning-based rapid entire body assessment (REBA) method has been proposed. The method takes working videos as input and automatically outputs the corresponding REBA score through 3D pose reconstruction. The proposed method achieves an average precision of 94.7% on real-world data, which is comparable to that of ergonomic experts. Furthermore, the method has the potential to be applied across a wide range of industries as it has demonstrated good generalisation in multiple scenarios. The proposed method offers a promising solution for automated and accurate risk assessment of WMSDs, with implications for various industries to ensure the safety and well-being of workers.

PMID:38423143 | DOI:10.1080/00140139.2024.2306315

Lancet Oncol. 2024 Mar;25(3):400-410. doi: 10.1016/S1470-2045(23)00641-1.

ABSTRACT

BACKGROUND: The extended acquisition times required for MRI limit its availability in resource-constrained settings. Consequently, accelerating MRI by undersampling k-space data, which is necessary to reconstruct an image, has been a long-standing but important challenge. We aimed to develop a deep convolutional neural network (dCNN) optimisation method for MRI reconstruction and to reduce scan times and evaluate its effect on image quality and accuracy of oncological imaging biomarkers.

METHODS: In this multicentre, retrospective, cohort study, MRI data from patients with glioblastoma treated at Heidelberg University Hospital (775 patients and 775 examinations) and from the phase 2 CORE trial (260 patients, 1083 examinations, and 58 institutions) and the phase 3 CENTRIC trial (505 patients, 3147 examinations, and 139 institutions) were used to develop, train, and test dCNN for reconstructing MRI from highly undersampled single-coil k-space data with various acceleration rates (R=2, 4, 6, 8, 10, and 15). Independent testing was performed with MRIs from the phase 2/3 EORTC-26101 trial (528 patients with glioblastoma, 1974 examinations, and 32 institutions). The similarity between undersampled dCNN-reconstructed and original MRIs was quantified with various image quality metrics, including structural similarity index measure (SSIM) and the accuracy of undersampled dCNN-reconstructed MRI on downstream radiological assessment of imaging biomarkers in oncology (automated artificial intelligence-based quantification of tumour burden and treatment response) was performed in the EORTC-26101 test dataset. The public NYU Langone Health fastMRI brain test dataset (558 patients and 558 examinations) was used to validate the generalisability and robustness of the dCNN for reconstructing MRIs from available multi-coil (parallel imaging) k-space data.

FINDINGS: In the EORTC-26101 test dataset, the median SSIM of undersampled dCNN-reconstructed MRI ranged from 0·88 to 0·99 across different acceleration rates, with 0·92 (95% CI 0·92-0·93) for 10-times acceleration (R=10). The 10-times undersampled dCNN-reconstructed MRI yielded excellent agreement with original MRI when assessing volumes of contrast-enhancing tumour (median DICE for spatial agreement of 0·89 [95% CI 0·88 to 0·89]; median volume difference of 0·01 cm3 [95% CI 0·00 to 0·03] equalling 0·21%; p=0·0036 for equivalence) or non-enhancing tumour or oedema (median DICE of 0·94 [95% CI 0·94 to 0·95]; median volume difference of -0·79 cm3 [95% CI -0·87 to -0·72] equalling -1·77%; p=0·023 for equivalence) in the EORTC-26101 test dataset. Automated volumetric tumour response assessment in the EORTC-26101 test dataset yielded an identical median time to progression of 4·27 months (95% CI 4·14 to 4·57) when using 10-times-undersampled dCNN-reconstructed or original MRI (log-rank p=0·80) and agreement in the time to progression in 374 (95·2%) of 393 patients with data. The dCNN generalised well to the fastMRI brain dataset, with significant improvements in the median SSIM when using multi-coil compared with single-coil k-space data (p<0·0001).

INTERPRETATION: Deep-learning-based reconstruction of undersampled MRI allows for a substantial reduction of scan times, with a 10-times acceleration demonstrating excellent image quality while preserving the accuracy of derived imaging biomarkers for the assessment of oncological treatment response. Our developments are available as open source software and hold considerable promise for increasing the accessibility to MRI, pending further prospective validation.

FUNDING: Deutsche Forschungsgemeinschaft (German Research Foundation) and an Else Kröner Clinician Scientist Endowed Professorship by the Else Kröner Fresenius Foundation.

PMID:38423052 | DOI:10.1016/S1470-2045(23)00641-1

Comput Biol Med. 2024 Feb 28;171:108212. doi: 10.1016/j.compbiomed.2024.108212. Online ahead of print.

ABSTRACT

BACKGROUND: Deep learning-based super-resolution (SR) algorithms aim to reconstruct low-resolution (LR) images into high-fidelity high-resolution (HR) images by learning the low- and high-frequency information. Experts' diagnostic requirements are fulfilled in medical application scenarios through the high-quality reconstruction of LR digital medical images.

PURPOSE: Medical image SR algorithms should satisfy the requirements of arbitrary resolution and high efficiency in applications. However, there is currently no relevant study available. Several SR research on natural images have accomplished the reconstruction of resolutions without limitations. However, these methodologies provide challenges in meeting medical applications due to the large scale of the model, which significantly limits efficiency. Hence, we suggest a highly effective method for reconstructing medical images at any desired resolution.

METHODS: Statistical features of medical images exhibit greater continuity in the region of neighboring pixels than natural images. Hence, the process of reconstructing medical images is comparatively less challenging. Utilizing this property, we develop a neighborhood evaluator to represent the continuity of the neighborhood while controlling the network's depth.

RESULTS: The suggested method has superior performance across seven scales of reconstruction, as evidenced by experiments conducted on panoramic radiographs and two external public datasets. Furthermore, the proposed network significantly decreases the parameter count by over 20× and the computational workload by over 10× compared to prior researches. On large-scale reconstruction, the inference speed can be enhanced by over 5×.

CONCLUSION: The novel proposed SR strategy for medical images performs efficient reconstruction at arbitrary resolution, marking a significant breakthrough in the field. The given scheme facilitates the implementation of SR in mobile medical platforms.

PMID:38422967 | DOI:10.1016/j.compbiomed.2024.108212

Comput Biol Med. 2024 Feb 28;171:108237. doi: 10.1016/j.compbiomed.2024.108237. Online ahead of print.

ABSTRACT

Lumbar vertebral body cancellous bone location and segmentation is crucial in an automated lumbar spine processing pipeline. Accurate and reliable analysis of lumbar spine image is expected to advantage practical medical diagnosis and population-based analysis of bone strength. However, the design of automated algorithms for lumbar spine processing is demanding due to significant anatomical variations and scarcity of publicly available data. In recent years, convolutional neural network (CNN) and vision transformers (Vits) have been the de facto standard in medical image segmentation. Although adept at capturing global features, the inherent bias of locality and weight sharing of CNN constrains its capacity to model long-range dependency. In contrast, Vits excel at long-range dependency modeling, but they may not generalize well with limited datasets due to the lack of inductive biases inherent to CNN. In this paper, we propose a deep learning-based two-stage coarse-to-fine solution to address the problem of automatic location and segmentation of lumbar vertebral body cancellous bone. Specifically, in the first stage, a Swin-transformer based model is applied to predict the heatmap of lumbar vertebral body centroids. Considering the characteristic anatomical structure of lumbar spine, we propose a novel loss function called LumAnatomy loss, which enforces the order and bend of the predicted vertebral body centroids. To inherit the excellence of CNN and Vits while preventing their respective limitations, in the second stage, we propose an encoder-decoder network to segment the identified lumbar vertebral body cancellous bone, which consists of two parallel encoders, i.e., a Swin-transformer encoder and a CNN encoder. To enhance the combination of CNNs and Vits, we propose a novel multi-scale attention feature fusion module (MSA-FFM), which address issues that arise when fusing features given at different encoders. To tackle the issue of lack of data, we raise the first large-scale lumbar vertebral body cancellous bone segmentation dataset called LumVBCanSeg containing a total of 185 CT scans annotated at voxel level by 3 physicians. Extensive experimental results on the LumVBCanSeg dataset demonstrate the proposed algorithm outperform other state-of-the-art medical image segmentation methods. The data is publicly available at: https://zenodo.org/record/8181250. The implementation of the proposed method is available at: https://github.com/sia405yd/LumVertCancNet.

PMID:38422966 | DOI:10.1016/j.compbiomed.2024.108237

Comput Biol Med. 2024 Feb 28;171:108228. doi: 10.1016/j.compbiomed.2024.108228. Online ahead of print.

ABSTRACT

Weakly supervised learning with image-level labels, releasing deep learning from highly labor-intensive pixel-wise annotation, has gained great attention for medical image segmentation. However, existing weakly supervised methods are mainly designed for single-class segmentation while leaving multi-class medical image segmentation rarely-explored. Different from natural images, label symbiosis, together with location adjacency, are much more common in medical images, making it more challenging for multi-class segmentation. In this paper, we propose a novel weakly supervised learning method for multi-class medical image segmentation with image-level labels. In terms of the multi-class classification backbone, a multi-level classification network encoding multi-scale features is proposed to produce binary predictions, together with the corresponding CAMs, of each class separately. To address the above issues (i.e., label symbiosis and location adjacency), a feature decomposition module based on semantic affinity is first proposed to learn both class-independent and class-dependent features by maximizing the inter-class feature distance. Through a cross-guidance loss to jointly utilize the above features, label symbiosis is largely alleviated. In terms of location adjacency, a mutually exclusive loss is constructed to minimize the overlap among regions corresponding to different classes. Experimental results on three datasets demonstrate the superior performance of the proposed weakly-supervised framework for both single-class and multi-class medical image segmentation. We believe the analysis in this paper would shed new light on future work for multi-class medical image segmentation. The source code of this paper is publicly available at https://github.com/HustAlexander/MCWSS.

PMID:38422964 | DOI:10.1016/j.compbiomed.2024.108228

Phys Med. 2024 Feb 28;119:103319. doi: 10.1016/j.ejmp.2024.103319. Online ahead of print.

ABSTRACT

PURPOSE: To characterise the impact of Precise Image (PI) deep learning reconstruction algorithm on image quality, compared to filtered back-projection (FBP) and iDose4 iterative reconstruction for brain computed tomography (CT) phantom images.

METHODS: Catphan-600 phantom was acquired with an Incisive CT scanner using a dedicated brain protocol, at six different dose levels (volume computed tomography dose index (CTDIvol): 7/14/29/49/56/67 mGy). Images were reconstructed using FBP, levels 2/5 of iDose4, and PI algorithm (Sharper/Sharp/Standard/Smooth/Smoother). Image quality was assessed by evaluating CT numbers, image histograms, noise, image non-uniformity (NU), noise power spectrum, target transfer function, and detectability index.

RESULTS: The five PI levels did not significantly affect the mean CT number. For a given CTDIvol using Sharper-to-Smoother levels, the spatial resolution for all the investigated materials and the detectability index increased while the noise magnitude decreased, slightly affecting noise texture. For a fixed PI level increasing the CTDIvol the detectability index increased, the noise magnitude decreased. From 29 mGy, NU values converged within 1 Hounsfield Unit from each other without a substantial improvement at higher CTDIvol values.

CONCLUSIONS: The improved performances of intermediate PI levels in brain protocols compared to conventional algorithms seem to suggest a potential reduction of CTDIvol.

PMID:38422902 | DOI:10.1016/j.ejmp.2024.103319

Phys Med Biol. 2024 Feb 29. doi: 10.1088/1361-6560/ad2ee7. Online ahead of print.

ABSTRACT

Concern has been expressed regarding the risk of carcinogenesis from medical computed tomography (CT) radiation. Lowering radiation in CT without appropriate modifications often leads to severe noise-induced artifacts in the images. The utilization of Deep Learning (DL) techniques has achieved promising reconstruction performance in low-dose CT (LDCT) imaging. However, most DL-based algorithms require the pre-collection of a large set of image pairs (low-dose/standard-dose) and the training of networks in an end-to-end supervised manner. Meanwhile, securing such a large volume of paired, well-registered training data in clinical practice is challenging. Moreover, these algorithms often overlook the potential to utilize the abundant information in a large collection of LDCT-only images/sinograms. In this paper, we introduce a Semi-supervised Iterative Adaptive Network (SIA-Net) for LDCT imaging, utilizing both labeled and unlabeled sinograms in a cohesive network framework, integrating supervised and unsupervised learning processes. Specifically, the supervised process captures critical features (i.e., noise distribution and tissue characteristics) latent in the paired sinograms, while the unsupervised process effectively learns these features in the unlabeled low-dose sinograms, employing a conventional weighted least-squares model with a regularization term. Furthermore, the SIA-Net method is designed to adaptively transfer the learned feature distribution from the supervised to the unsupervised process, thereby obtaining a high-fidelity sinogram through iterative adaptive learning. Finally, high-quality CT images can be reconstructed from the refined sinogram using the filtered back-projection algorithm. Experimental results on two clinical datasets indicate that the proposed SIA-Net method achieves competitive performance in terms of noise reduction and structure preservation in LDCT imaging, when compared to traditional supervised learning methods.

PMID:38422540 | DOI:10.1088/1361-6560/ad2ee7

PLoS Comput Biol. 2024 Feb 29;20(2):e1011774. doi: 10.1371/journal.pcbi.1011774. eCollection 2024 Feb.

ABSTRACT

Dendritic spines are the seat of most excitatory synapses in the brain, and a cellular structure considered central to learning, memory, and activity-dependent plasticity. The quantification of dendritic spines from light microscopy data is usually performed by humans in a painstaking and error-prone process. We found that human-to-human variability is substantial (inter-rater reliability 82.2±6.4%), raising concerns about the reproducibility of experiments and the validity of using human-annotated 'ground truth' as an evaluation method for computational approaches of spine identification. To address this, we present DeepD3, an open deep learning-based framework to robustly quantify dendritic spines in microscopy data in a fully automated fashion. DeepD3's neural networks have been trained on data from different sources and experimental conditions, annotated and segmented by multiple experts and they offer precise quantification of dendrites and dendritic spines. Importantly, these networks were validated in a number of datasets on varying acquisition modalities, species, anatomical locations and fluorescent indicators. The entire DeepD3 open framework, including the fully segmented training data, a benchmark that multiple experts have annotated, and the DeepD3 model zoo is fully available, addressing the lack of openly available datasets of dendritic spines while offering a ready-to-use, flexible, transparent, and reproducible spine quantification method.

PMID:38422112 | DOI:10.1371/journal.pcbi.1011774

PLoS One. 2024 Feb 29;19(2):e0299366. doi: 10.1371/journal.pone.0299366. eCollection 2024.

ABSTRACT

PURPOSE: To conduct a volumetric and movement analysis of lung parenchyma in prone positioning using deep neural networks (DNNs).

METHOD: We included patients with suspected interstitial lung abnormalities or disease who underwent full-inspiratory supine and prone chest CT at a single institution between June 2021 and March 2022. A thoracic radiologist visually assessed the fibrosis extent in the total lung (using units of 10%) on supine CT. After preprocessing the images into 192×192×192 resolution, a DNN automatically segmented the whole lung and pulmonary lobes in prone and supine CT images. Affine registration matched the patient's center and location, and the DNN deformably registered prone and supine CT images to calculate the x-, y-, z-axis, and 3D pixel movements.

RESULTS: In total, 108 CT pairs had successful registration. Prone positioning significantly increased the left lower (90.2±69.5 mL, P = 0.000) and right lower lobar volumes (52.5±74.2 mL, P = 0.000). During deformable registration, the average maximum whole-lung pixel movements between the two positions were 1.5, 1.9, 1.6, and 2.8 cm in each axis and 3D plane. Compared to patients with <30% fibrosis, those with ≥30% fibrosis had smaller volume changes (P<0.001) and smaller pixel movements in all axes between the positions (P = 0.000-0.007). Forced vital capacity (FVC) correlated with the left lower lobar volume increase (Spearman correlation coefficient, 0.238) and the maximum whole-lung pixel movements in all axes (coefficients, 0.311 to 0.357).

CONCLUSIONS: Prone positioning led to the preferential expansion of the lower lobes, correlated with FVC, and lung fibrosis limited lung expansion during prone positioning.

PMID:38422097 | DOI:10.1371/journal.pone.0299366

PLoS One. 2024 Feb 29;19(2):e0298228. doi: 10.1371/journal.pone.0298228. eCollection 2024.

ABSTRACT

This article investigates whether higher classification accuracy can always be achieved by utilizing a pre-trained deep learning model as the feature extractor in the Bag-of-Deep-Visual-Words (BoDVW) classification model, as opposed to directly using the new classification layer of the pre-trained model for classification. Considering the multiple factors related to the feature extractor -such as model architecture, fine-tuning strategy, number of training samples, feature extraction method, and feature encoding method-we investigate these factors through experiments and then provide detailed answers to the question. In our experiments, we use five feature encoding methods: hard-voting, soft-voting, locally constrained linear coding, super vector coding, and fisher vector (FV). We also employ two popular feature extraction methods: one (denoted as Ext-DFs(CP)) uses a convolutional or non-global pooling layer, and another (denoted as Ext-DFs(FC)) uses a fully-connected or global pooling layer. Three pre-trained models-VGGNet-16, ResNext-50(32×4d), and Swin-B-are utilized as feature extractors. Experimental results on six datasets (15-Scenes, TF-Flowers, MIT Indoor-67, COVID-19 CXR, NWPU-RESISC45, and Caltech-101) reveal that compared to using the pre-trained model with only the new classification layer re-trained for classification, employing it as the feature extractor in the BoDVW model improves the accuracy in 35 out of 36 experiments when using FV. With Ext-DFs(CP), the accuracy increases by 0.13% to 8.43% (averaged at 3.11%), and with Ext-DFs(FC), it increases by 1.06% to 14.63% (averaged at 5.66%). Furthermore, when all layers of the pre-trained model are fine-tuned and used as the feature extractor, the results vary depending on the methods used. If FV and Ext-DFs(FC) are used, the accuracy increases by 0.21% to 5.65% (averaged at 1.58%) in 14 out of 18 experiments. Our results suggest that while using a pre-trained deep learning model as the feature extractor does not always improve classification accuracy, it holds great potential as an accuracy improvement technique.

PMID:38422007 | DOI:10.1371/journal.pone.0298228

PLoS One. 2024 Feb 29;19(2):e0297793. doi: 10.1371/journal.pone.0297793. eCollection 2024.

ABSTRACT

Prediction of major arrhythmic events (MAEs) in dilated cardiomyopathy represents an unmet clinical goal. Computational models and artificial intelligence (AI) are new technological tools that could offer a significant improvement in our ability to predict MAEs. In this proof-of-concept study, we propose a deep learning (DL)-based model, which we termed Deep ARrhythmic Prevention in dilated cardiomyopathy (DARP-D), built using multidimensional cardiac magnetic resonance data (cine videos and hypervideos and LGE images and hyperimages) and clinical covariates, aimed at predicting and tracking an individual patient's risk curve of MAEs (including sudden cardiac death, cardiac arrest due to ventricular fibrillation, sustained ventricular tachycardia lasting ≥30 s or causing haemodynamic collapse in <30 s, appropriate implantable cardiac defibrillator intervention) over time. The model was trained and validated in 70% of a sample of 154 patients with dilated cardiomyopathy and tested in the remaining 30%. DARP-D achieved a 95% CI in Harrell's C concordance indices of 0.12-0.68 on the test set. We demonstrate that our DL approach is feasible and represents a novelty in the field of arrhythmic risk prediction in dilated cardiomyopathy, able to analyze cardiac motion, tissue characteristics, and baseline covariates to predict an individual patient's risk curve of major arrhythmic events. However, the low number of patients, MAEs and epoch of training make the model a promising prototype but not ready for clinical usage. Further research is needed to improve, stabilize and validate the performance of the DARP-D to convert it from an AI experiment to a daily used tool.

PMID:38421987 | DOI:10.1371/journal.pone.0297793

IEEE Trans Pattern Anal Mach Intell. 2024 Feb 29;PP. doi: 10.1109/TPAMI.2024.3370716. Online ahead of print.

ABSTRACT

Randomness is widely introduced in neural network training to simplify model optimization or avoid the over-fitting problem. Among them, dropout and its variations in different aspects (e.g., data, model structure) are prevalent in regularizing the training of deep neural networks. Though effective and performing well, the randomness introduced by these dropout-based methods causes nonnegligible inconsistency between training and inference. In this paper, we introduce a simple consistency training strategy to regularize such randomness, namely R-Drop, which forces two output distributions sampled by each type of randomness to be consistent. Specifically, R-Drop minimizes the bidirectional KL-divergence between two output distributions produced by dropout-based randomness for each training sample. Theoretical analysis reveals that R-Drop can reduce the above inconsistency by reducing the inconsistency among the sampled sub structures and bridging the gap between the loss calculated by the full model and sub structures. Experiments on 7 widely-used deep learning tasks ( 23 datasets in total) demonstrate that R-Drop is universally effective for different types of neural networks (i.e., feed-forward, recurrent, and graph neural networks) and different learning paradigms (supervised, parameter-efficient, and semi-supervised). In particular, it achieves state-of-the-art performances with the vanilla Transformer model on WMT14 English → German translation ( 30.91 BLEU) and WMT14 English → French translation ( 43.95 BLEU), even surpassing models trained with extra large-scale data and expert-designed advanced variants of Transformer models. Our code is available at GitHub https://github.com/dropreg/R-Drop.

PMID:38421846 | DOI:10.1109/TPAMI.2024.3370716

IEEE Trans Neural Syst Rehabil Eng. 2024 Feb 29;PP. doi: 10.1109/TNSRE.2024.3371896. Online ahead of print.

ABSTRACT

The development of advanced prosthetic devices that can be seamlessly used during an individual's daily life remains a significant challenge in the field of rehabilitation engineering. This study compares the performance of deep learning architectures to shallow networks in decoding motor intent for prosthetic control using electromyography (EMG) signals. Four neural network architectures, including a feedforward neural network with one hidden layer, a feedforward neural network with multiple hidden layers, a temporal convolutional network, and a convolutional neural network with squeeze-and-excitation operations were evaluated in real-time, human-in-the-loop experiments with able-bodied participants and an individual with an amputation. Our results demonstrate that deep learning architectures outperform shallow networks in decoding motor intent, with representation learning effectively extracting underlying motor control information from EMG signals. Furthermore, the observed performance improvements by using deep neural networks were consistent across both able-bodied and amputee participants. By employing deep neural networks instead of a shallow network, more reliable and precise control of a prosthesis can be achieved, which has the potential to significantly enhance prosthetic functionality and improve the quality of life for individuals with amputations.

PMID:38421839 | DOI:10.1109/TNSRE.2024.3371896

Cytometry A. 2024 Feb 29. doi: 10.1002/cyto.a.24832. Online ahead of print.

ABSTRACT

The gold standard of leukocyte differentiation is a manual examination of blood smears, which is not only time and labor intensive but also susceptible to human error. As to automatic classification, there is still no comparative study of cell segmentation, feature extraction, and cell classification, where a variety of machine and deep learning models are compared with home-developed approaches. In this study, both traditional machine learning of K-means clustering versus deep learning of U-Net, U-Net + ResNet18, and U-Net + ResNet34 were used for cell segmentation, producing segmentation accuracies of 94.36% versus 99.17% for the dataset of CellaVision and 93.20% versus 98.75% for the dataset of BCCD, confirming that deep learning produces higher performance than traditional machine learning in leukocyte classification. In addition, a series of deep-learning approaches, including AlexNet, VGG16, and ResNet18, was adopted to conduct feature extraction and cell classification of leukocytes, producing classification accuracies of 91.31%, 97.83%, and 100% of CellaVision as well as 81.18%, 91.64% and 97.82% of BCCD, confirming the capability of the increased deepness of neural networks in leukocyte classification. As to the demonstrations, this study further conducted cell-type classification of ALL-IDB2 and PCB-HBC datasets, producing high accuracies of 100% and 98.49% among all literature, validating the deep learning model used in this study.

PMID:38420862 | DOI:10.1002/cyto.a.24832

Front Biosci (Landmark Ed). 2024 Feb 21;29(2):75. doi: 10.31083/j.fbl2902075.

ABSTRACT

BACKGROUND: Cerebral Cavernous Malformations (CCMs) are brain vascular abnormalities associated with an increased risk of hemorrhagic strokes. Familial CCMs result from autosomal dominant inheritance involving three genes: KRIT1 (CCM1), MGC4607 (CCM2), and PDCD10 (CCM3). CCM1 and CCM3 form the CCM Signal Complex (CSC) by binding to CCM2. Both CCM1 and CCM2 exhibit cellular heterogeneity through multiple alternative spliced isoforms, where exons from the same gene combine in diverse ways, leading to varied mRNA transcripts. Additionally, both demonstrate nucleocytoplasmic shuttling between the nucleus and cytoplasm, suggesting their potential role in gene expression regulation as transcription factors (TFs). Due to the accumulated data indicating the cellular localization of CSC proteins in the nucleus and their interaction with progesterone receptors, which serve dual roles as both cellular signaling components and TFs, a question has arisen regarding whether CCMs could also function in both capacities like progesterone receptors.

METHODS: To investigate this potential, we employed our proprietary deep-learning (DL)-based algorithm, specifically utilizing a biased-Support Vector Machine (SVM) model, to explore the plausible cellular function of any of the CSC proteins, particularly focusing on CCM gene isoforms with nucleocytoplasmic shuttling, acting as TFs in gene expression regulation.

RESULTS: Through a comparative DL-based predictive analysis, we have effectively discerned a collective of 11 isoforms across all CCM proteins (CCM1-3). Additionally, we have substantiated the TF functionality of 8 isoforms derived from CCM1 and CCM2 proteins, marking the inaugural identification of CCM isoforms in the role of TFs.

CONCLUSIONS: This groundbreaking discovery directly challenges the prevailing paradigm, which predominantly emphasizes the involvement of CSC solely in endothelial cellular functions amid various potential cellular signal cascades during angiogenesis.

PMID:38420834 | DOI:10.31083/j.fbl2902075