Curr Cardiol Rep. 2024 May 16. doi: 10.1007/s11886-024-02062-1. Online ahead of print.

ABSTRACT

PURPOSE OF REVIEW: Artificial intelligence (AI) is transforming electrocardiography (ECG) interpretation. AI diagnostics can reach beyond human capabilities, facilitate automated access to nuanced ECG interpretation, and expand the scope of cardiovascular screening in the population. AI can be applied to the standard 12-lead resting ECG and single-lead ECGs in external monitors, implantable devices, and direct-to-consumer smart devices. We summarize the current state of the literature on AI-ECG.

RECENT FINDINGS: Rhythm classification was the first application of AI-ECG. Subsequently, AI-ECG models have been developed for screening structural heart disease including hypertrophic cardiomyopathy, cardiac amyloidosis, aortic stenosis, pulmonary hypertension, and left ventricular systolic dysfunction. Further, AI models can predict future events like development of systolic heart failure and atrial fibrillation. AI-ECG exhibits potential in acute cardiac events and non-cardiac applications, including acute pulmonary embolism, electrolyte abnormalities, monitoring drugs therapy, sleep apnea, and predicting all-cause mortality. Many AI models in the domain of cardiac monitors and smart watches have received Food and Drug Administration (FDA) clearance for rhythm classification, while others for identification of cardiac amyloidosis, pulmonary hypertension and left ventricular dysfunction have received breakthrough device designation. As AI-ECG models continue to be developed, in addition to regulatory oversight and monetization challenges, thoughtful clinical implementation to streamline workflows, avoiding information overload and overwhelming of healthcare systems with false positive results is necessary. Research to demonstrate and validate improvement in healthcare efficiency and improved patient outcomes would be required before widespread adoption of any AI-ECG model.

PMID:38753291 | DOI:10.1007/s11886-024-02062-1

Eur Radiol. 2024 May 16. doi: 10.1007/s00330-024-10793-6. Online ahead of print.

ABSTRACT

OBJECTIVES: To investigate the feasibility of low-radiation dose and low iodinated contrast medium (ICM) dose protocol combining low-tube voltage and deep-learning reconstruction (DLR) algorithm in thin-slice abdominal CT.

METHODS: This prospective study included 148 patients who underwent contrast-enhanced abdominal CT with either 120-kVp (600 mgL/kg, n = 74) or 80-kVp protocol (360 mgL/kg, n = 74). The 120-kVp images were reconstructed using hybrid iterative reconstruction (HIR) (120-kVp-HIR), while 80-kVp images were reconstructed using HIR (80-kVp-HIR) and DLR (80-kVp-DLR) with 0.5 mm thickness. Size-specific dose estimate (SSDE) and iodine dose were compared between protocols. Image noise, CT attenuation, and contrast-to-noise ratio (CNR) were quantified. Noise power spectrum (NPS) and edge rise slope (ERS) were used to evaluate noise texture and edge sharpness, respectively. The subjective image quality was rated on a 4-point scale.

RESULTS: SSDE and iodine doses of 80-kVp were 40.4% (8.1 ± 0.9 vs. 13.6 ± 2.7 mGy) and 36.3% (21.2 ± 3.9 vs. 33.3 ± 4.3 gL) lower, respectively, than those of 120-kVp (both, p < 0.001). CT attenuation of vessels and solid organs was higher in 80-kVp than in 120-kVp images (all, p < 0.001). Image noise of 80-kVp-HIR and 80-kVp-DLR was higher and lower, respectively than that of 120-kVp-HIR (both p < 0.001). The highest CNR and subjective scores were attained in 80-kVp-DLR (all, p < 0.001). There were no significant differences in average NPS frequency and ERS between 120-kVp-HIR and 80-kVp-DLR (p ≥ 0.38).

CONCLUSION: Compared with the 120-kVp-HIR protocol, the combined use of 80-kVp and DLR techniques yielded superior subjective and objective image quality with reduced radiation and ICM doses at thin-section abdominal CT.

CLINICAL RELEVANCE STATEMENT: Scanning at low-tube voltage (80-kVp) combined with the deep-learning reconstruction algorithm may enhance diagnostic efficiency and patient safety by improving image quality and reducing radiation and contrast doses of thin-slice abdominal CT.

KEY POINTS: Reducing radiation and iodine doses is desirable; however, contrast and noise degradation can be detrimental. The 80-kVp scan with the deep-learning reconstruction technique provided better images with lower radiation and contrast doses. This technique may be efficient for improving diagnostic confidence and patient safety in thin-slice abdominal CT.

PMID:38753193 | DOI:10.1007/s00330-024-10793-6

Network. 2024 May 16:1-32. doi: 10.1080/0954898X.2024.2350579. Online ahead of print.

ABSTRACT

One of the most used diagnostic imaging techniques for identifying a variety of lung and bone-related conditions is the chest X-ray. Recent developments in deep learning have demonstrated several successful cases of illness diagnosis from chest X-rays. However, issues of stability and class imbalance still need to be resolved. Hence in this manuscript, multi-class lung disease classification in chest x-ray images using a hybrid manta-ray foraging volcano eruption algorithm boosted multilayer perceptron neural network approach is proposed (MPNN-Hyb-MRF-VEA). Initially, the input chest X-ray images are taken from the Covid-Chest X-ray dataset. Anisotropic diffusion Kuwahara filtering (ADKF) is used to enhance the quality of these images and lower noise. To capture significant discriminative features, the Term frequency-inverse document frequency (TF-IDF) based feature extraction method is utilized in this case. The Multilayer Perceptron Neural Network (MPNN) serves as the classification model for multi-class lung disorders classification as COVID-19, pneumonia, tuberculosis (TB), and normal. A Hybrid Manta-Ray Foraging and Volcano Eruption Algorithm (Hyb-MRF-VEA) is introduced to further optimize and fine-tune the MPNN's parameters. The Python platform is used to accurately evaluate the proposed methodology. The performance of the proposed method provides 23.21%, 12.09%, and 5.66% higher accuracy compared with existing methods like NFM, SVM, and CNN respectively.

PMID:38753162 | DOI:10.1080/0954898X.2024.2350579

Int J Clin Pharm. 2024 May 16. doi: 10.1007/s11096-024-01724-y. Online ahead of print.

ABSTRACT

BACKGROUND: Venlafaxine is frequently prescribed for patients with depression. To control the concentration of venlafaxine within the therapeutic window for the best treatment effect, a model to predict venlafaxine concentration is necessary.

AIM: Our objective was to develop a prediction model for venlafaxine concentration using real-world evidence based on machine learning and deep learning techniques.

METHOD: Patients who underwent venlafaxine treatment between November 2019 and August 2022 were included in the study. Important variables affecting venlafaxine concentration were identified using a combination of univariate analysis, sequential forward selection, and machine learning techniques. Predictive performance of nine machine learning and deep learning algorithms were assessed, and the one with the optimal performance was selected for modeling. The final model was interpreted using SHapley Additive exPlanations.

RESULTS: A total of 330 eligible patients were included. Five influential variables that affect venlafaxine concentration were venlafaxine daily dose, sex, age, hyperlipidemia, and adenosine deaminase. The venlafaxine concentration prediction model was developed using the eXtreme Gradient Boosting algorithm (R2 = 0.65, mean absolute error = 77.92, root mean square error = 93.58). In the testing cohort, the accuracy of the predicted concentration within ± 30% of the actual concentration was 73.49%. In the subgroup analysis, the prediction accuracy was 69.39% within the recommended therapeutic range of venlafaxine concentration within ± 30% of the actual value.

CONCLUSION: The XGBoost model for predicting blood concentration of venlafaxine using real-world evidence was developed, guiding the adjustment of regimen in clinical practice.

PMID:38753076 | DOI:10.1007/s11096-024-01724-y

Chemphyschem. 2024 May 16:e202400176. doi: 10.1002/cphc.202400176. Online ahead of print.

ABSTRACT

We report a deep learning-based approach to accurately predict the emission spectra of phosphorescent heteroleptic [Ir(C^N)2(NN)]+ complexes, enabling the rapid discovery of novel Ir(III) chromophores for diverse applications including organic light-emitting diodes and solar fuel cells. The deep learning models utilize graph neural networks and other chemical features in architectures that reflect the inherent structure of the heteroleptic complexes, composed of C^N and N^N ligands, and are thus geared towards efficient training over the dataset. By leveraging experimental emission data, our models reliably predict the full emission spectra of these complexes across various emission profiles, surpassing the accuracy of conventional DFT and correlated wavefunction methods, while simultaneously achieving robustness to the presence of imperfect (noisy, low-quality) training spectra. We showcase the potential applications for these and related models for \insilico\ prediction of complexes with tailored emission properties, as well as in "design of experiment'' contexts to reduce the synthetic burden of high-throughput screening. In the latter case, we demonstrate that the models allow to exploit a limited amount of experimental data to explore a wide range of chemical space, thus leveraging a modest synthetic effort.

PMID:38752882 | DOI:10.1002/cphc.202400176

Brief Bioinform. 2024 Mar 27;25(3):bbae224. doi: 10.1093/bib/bbae224.

ABSTRACT

Avian reoviruses continue to cause disease in turkeys with varied pathogenicity and tissue tropism. Turkey enteric reovirus has been identified as a causative agent of enteritis or inapparent infections in turkeys. The new emerging variants of turkey reovirus, tentatively named turkey arthritis reovirus (TARV) and turkey hepatitis reovirus (THRV), are linked to tenosynovitis/arthritis and hepatitis, respectively. Turkey arthritis and hepatitis reoviruses are causing significant economic losses to the turkey industry. These infections can lead to poor weight gain, uneven growth, poor feed conversion, increased morbidity and mortality and reduced marketability of commercial turkeys. To combat these issues, detecting and classifying the types of reoviruses in turkey populations is essential. This research aims to employ clustering methods, specifically K-means and Hierarchical clustering, to differentiate three types of turkey reoviruses and identify novel emerging variants. Additionally, it focuses on classifying variants of turkey reoviruses by leveraging various machine learning algorithms such as Support Vector Machines, Naive Bayes, Random Forest, Decision Tree, and deep learning algorithms, including convolutional neural networks (CNNs). The experiments use real turkey reovirus sequence data, allowing for robust analysis and evaluation of the proposed methods. The results indicate that machine learning methods achieve an average accuracy of 92%, F1-Macro of 93% and F1-Weighted of 92% scores in classifying reovirus types. In contrast, the CNN model demonstrates an average accuracy of 85%, F1-Macro of 71% and F1-Weighted of 84% scores in the same classification task. The superior performance of the machine learning classifiers provides valuable insights into reovirus evolution and mutation, aiding in detecting emerging variants of pathogenic TARVs and THRVs.

PMID:38752857 | DOI:10.1093/bib/bbae224

Crit Care Med. 2024 Jun 1;52(6):e330-e331. doi: 10.1097/CCM.0000000000006291. Epub 2024 May 16.

NO ABSTRACT

PMID:38752829 | DOI:10.1097/CCM.0000000000006291

Technol Cancer Res Treat. 2024 Jan-Dec;23:15330338241254060. doi: 10.1177/15330338241254060.

ABSTRACT

Objectives: This study aimed to build a comprehensive deep-learning model for the prediction of radiation pneumonitis using chest computed tomography (CT), clinical, dosimetric, and laboratory data. Introduction: Radiation therapy is an effective tool for treating patients with lung cancer. Despite its effectiveness, the risk of radiation pneumonitis limits its application. Although several studies have demonstrated models to predict radiation pneumonitis, no reliable model has been developed yet. Herein, we developed prediction models using pretreatment chest CT and various clinical data to assess the likelihood of radiation pneumonitis in lung cancer patients. Methods: This retrospective study analyzed 3-dimensional (3D) lung volume data from chest CT scans and 27 features including dosimetric, clinical, and laboratory data from 548 patients who were treated at our institution between 2010 and 2021. We developed a neural network, named MergeNet, which processes lung 3D CT, clinical, dosimetric, and laboratory data. The MergeNet integrates a convolutional neural network with subsequent fully connected layers. A support vector machine (SVM) and light gradient boosting machine (LGBM) model were also implemented for comparison. For comparison, the convolution-only neural network was implemented as well. Three-dimensional Resnet-10 network and 4-fold cross-validation were used. Results: Classification performance was quantified by using the area under the receiver operative characteristic curve (AUC) metrics. MergeNet showed the AUC of 0.689. SVM, LGBM, and convolution-only networks showed AUCs of 0.525, 0.541, and 0.550, respectively. Application of DeLong test to pairs of receiver operating characteristic curves respectively yielded P values of .001 for the MergeNet-SVM pair and 0.001 for the MergeNet-LGBM pair. Conclusion: The MergeNet model, which incorporates chest CT, clinical, dosimetric, and laboratory data, demonstrated superior performance compared to other models. However, since its prediction performance has not yet reached an efficient level for clinical application, further research is required. Contribution: This study showed that MergeNet may be an effective means to predict radiation pneumonitis. Various predictive factors can be used together for the radiation pneumonitis prediction task via the MergeNet.

PMID:38752262 | DOI:10.1177/15330338241254060

Proc SPIE Int Soc Opt Eng. 2024 Feb;12933:129330T. doi: 10.1117/12.3006796. Epub 2024 Apr 3.

ABSTRACT

Tumor budding refers to a cluster of one to four tumor cells located at the tumor-invasive front. While tumor budding is a prognostic factor for colorectal cancer, counting and grading tumor budding are time consuming and not highly reproducible. There could be high inter- and intra-reader disagreement on H&E evaluation. This leads to the noisy training (imperfect ground truth) of deep learning algorithms, resulting in high variability and losing their ability to generalize on unseen datasets. Pan-cytokeratin staining is one of the potential solutions to enhance the agreement, but it is not routinely used to identify tumor buds and can lead to false positives. Therefore, we aim to develop a weakly-supervised deep learning method for tumor bud detection from routine H&E-stained images that does not require strict tissue-level annotations. We also propose Bayesian Multiple Instance Learning (BMIL) that combines multiple annotated regions during the training process to further enhance the generalizability and stability in tumor bud detection. Our dataset consists of 29 colorectal cancer H&E-stained images that contain 115 tumor buds per slide on average. In six-fold cross-validation, our method demonstrated an average precision and recall of 0.94, and 0.86 respectively. These results provide preliminary evidence of the feasibility of our approach in improving the generalizability in tumor budding detection using H&E images while avoiding the need for non-routine immunohistochemical staining methods.

PMID:38752165 | PMC:PMC11095418 | DOI:10.1117/12.3006796

Med Image Comput Comput Assist Interv. 2023 Oct;14220:651-662. doi: 10.1007/978-3-031-43907-0_62. Epub 2023 Oct 1.

ABSTRACT

Deep learning nowadays offers expert-level and sometimes even super-expert-level performance, but achieving such performance demands massive annotated data for training (e.g., Google's proprietary CXR Foundation Model (CXR-FM) was trained on 821,544 labeled and mostly private chest X-rays (CXRs)). Numerous datasets are publicly available in medical imaging but individually small and heterogeneous in expert labels. We envision a powerful and robust foundation model that can be trained by aggregating numerous small public datasets. To realize this vision, we have developed Ark, a framework that accrues and reuses knowledge from heterogeneous expert annotations in various datasets. As a proof of concept, we have trained two Ark models on 335,484 and 704,363 CXRs, respectively, by merging several datasets including ChestX-ray14, CheXpert, MIMIC-II, and VinDr-CXR, evaluated them on a wide range of imaging tasks covering both classification and segmentation via fine-tuning, linear-probing, and gender-bias analysis, and demonstrated our Ark's superior and robust performance over the state-of-the-art (SOTA) fully/self-supervised baselines and Google's proprietary CXR-FM. This enhanced performance is attributed to our simple yet powerful observation that aggregating numerous public datasets diversifies patient populations and accrues knowledge from diverse experts, yielding unprecedented performance yet saving annotation cost. With all codes and pretrained models released at GitHub.com/JLiangLab/Ark, we hope that Ark exerts an important impact on open science, as accruing and reusing knowledge from expert annotations in public datasets can potentially surpass the performance of proprietary models trained on unusually large data, inspiring many more researchers worldwide to share codes and datasets to build open foundation models, accelerate open science, and democratize deep learning for medical imaging.

PMID:38751905 | PMC:PMC11095392 | DOI:10.1007/978-3-031-43907-0_62

Front Plant Sci. 2024 May 1;15:1361309. doi: 10.3389/fpls.2024.1361309. eCollection 2024.

ABSTRACT

The identification of sweet corn seed vitality is an essential criterion for selecting high-quality varieties. In this research, a combination of hyperspectral imaging technique and diverse deep learning algorithms has been utilized to identify different vitality grades of sweet corn seeds. First, the hyperspectral data of 496 seeds, including four viability-grade seeds, are extracted and preprocessed. Then, support vector machine (SVM) and extreme learning machine (ELM) are used to construct the classification models. Finally, the one-dimensional convolutional neural networks (1DCNN), one-dimensional long short-term memory (1DLSTM), the CNN combined with the LSTM (CNN-LSTM), and the proposed firefly algorithm (FA) optimized CNN-LSTM (FA-CNN-LSTM) are utilized to distinguish spectral images of sweet corn seeds viability grade. The findings from the experimental analysis indicate that the deep learning models exhibit a significant advantage over traditional machine learning approaches in the discrimination of seed vitality levels, boasting a classification accuracy exceeding 94.26% in test datasets and achieving an accuracy improvement of at least 3% compared to the best-performing machine learning model. Moreover, the performance of the FA-CNN-LSTM model proposed in this study demonstrated a slight superiority over the other three models. Besides, the FA-CNN-LSTM achieved a classification accuracy of 97.23%, representing a significant improvement of 2.97% compared to the lowest-performing CNN and a 1.49% enhancement over the CNN-LSTM. In summary, this study reveals the potential of integrating deep learning with hyperspectral imaging as a promising alternative for discriminating sweet corn seed vitality grade, showcasing its value in agricultural research and cultivar breeding.

PMID:38751847 | PMC:PMC11094355 | DOI:10.3389/fpls.2024.1361309

Front Plant Sci. 2024 May 1;15:1399753. doi: 10.3389/fpls.2024.1399753. eCollection 2024.

NO ABSTRACT

PMID:38751845 | PMC:PMC11094301 | DOI:10.3389/fpls.2024.1399753

Front Plant Sci. 2024 May 1;15:1398903. doi: 10.3389/fpls.2024.1398903. eCollection 2024.

ABSTRACT

Sugarcane smut and Pachymetra root rots are two serious diseases of sugarcane, with susceptible infected crops losing over 30% of yield. A heritable component to both diseases has been demonstrated, suggesting selection could improve disease resistance. Genomic selection could accelerate gains even further, enabling early selection of resistant seedlings for breeding and clonal propagation. In this study we evaluated four types of algorithms for genomic predictions of clonal performance for disease resistance. These algorithms were: Genomic best linear unbiased prediction (GBLUP), including extensions to model dominance and epistasis, Bayesian methods including BayesC and BayesR, Machine learning methods including random forest, multilayer perceptron (MLP), modified convolutional neural network (CNN) and attention networks designed to capture epistasis across the genome-wide markers. Simple hybrid methods, that first used BayesR/GWAS to identify a subset of 1000 markers with moderate to large marginal additive effects, then used attention networks to derive predictions from these effects and their interactions, were also developed and evaluated. The hypothesis for this approach was that using a subset of markers more likely to have an effect would enable better estimation of interaction effects than when there were an extremely large number of possible interactions, especially with our limited data set size. To evaluate the methods, we applied both random five-fold cross-validation and a structured PCA based cross-validation that separated 4702 sugarcane clones (that had disease phenotypes and genotyped for 26k genome wide SNP markers) by genomic relationship. The Bayesian methods (BayesR and BayesC) gave the highest accuracy of prediction, followed closely by hybrid methods with attention networks. The hybrid methods with attention networks gave the lowest variation in accuracy of prediction across validation folds (and lowest MSE), which may be a criteria worth considering in practical breeding programs. This suggests that hybrid methods incorporating the attention mechanism could be useful for genomic prediction of clonal performance, particularly where non-additive effects may be important.

PMID:38751840 | PMC:PMC11095127 | DOI:10.3389/fpls.2024.1398903

Erkenntnis. 2024;89(5):1823-1840. doi: 10.1007/s10670-022-00605-y. Epub 2022 Aug 7.

ABSTRACT

Some machine learning models, in particular deep neural networks (DNNs), are not very well understood; nevertheless, they are frequently used in science. Does this lack of understanding pose a problem for using DNNs to understand empirical phenomena? Emily Sullivan has recently argued that understanding with DNNs is not limited by our lack of understanding of DNNs themselves. In the present paper, we will argue, contra Sullivan, that our current lack of understanding of DNNs does limit our ability to understand with DNNs. Sullivan's claim hinges on which notion of understanding is at play. If we employ a weak notion of understanding, then her claim is tenable, but rather weak. If, however, we employ a strong notion of understanding, particularly explanatory understanding, then her claim is not tenable.

PMID:38751773 | PMC:PMC11090801 | DOI:10.1007/s10670-022-00605-y

Adv Neural Inf Process Syst. 2023 Dec;36:33365-33378.

ABSTRACT

Transformers are widely used deep learning architectures. Existing transformers are mostly designed for sequences (texts or time series), images or videos, and graphs. This paper proposes a novel transformer model for massive (up to a million) point samples in continuous space. Such data are ubiquitous in environment sciences (e.g., sensor observations), numerical simulations (e.g., particle-laden flow, astrophysics), and location-based services (e.g., POIs and trajectories). However, designing a transformer for massive spatial points is non-trivial due to several challenges, including implicit long-range and multi-scale dependency on irregular points in continuous space, a non-uniform point distribution, the potential high computational costs of calculating all-pair attention across massive points, and the risks of over-confident predictions due to varying point density. To address these challenges, we propose a new hierarchical spatial transformer model, which includes multi-resolution representation learning within a quad-tree hierarchy and efficient spatial attention via coarse approximation. We also design an uncertainty quantification branch to estimate prediction confidence related to input feature noise and point sparsity. We provide a theoretical analysis of computational time complexity and memory costs. Extensive experiments on both real-world and synthetic datasets show that our method outperforms multiple baselines in prediction accuracy and our model can scale up to one million points on one NVIDIA A100 GPU. The code is available at https://github.com/spatialdatasciencegroup/HST.

PMID:38751689 | PMC:PMC11094554

ACS EST Air. 2024 Mar 27;1(5):332-345. doi: 10.1021/acsestair.3c00054. eCollection 2024 May 10.

ABSTRACT

Global fine particulate matter (PM2.5) assessment is impeded by a paucity of monitors. We improve estimation of the global distribution of PM2.5 concentrations by developing, optimizing, and applying a convolutional neural network with information from satellite-, simulation-, and monitor-based sources to predict the local bias in monthly geophysical a priori PM2.5 concentrations over 1998-2019. We develop a loss function that incorporates geophysical a priori estimates and apply it in model training to address the unrealistic results produced by mean-square-error loss functions in regions with few monitors. We introduce novel spatial cross-validation for air quality to examine the importance of considering spatial properties. We address the sharp decline in deep learning model performance in regions distant from monitors by incorporating the geophysical a priori PM2.5. The resultant monthly PM2.5 estimates are highly consistent with spatial cross-validation PM2.5 concentrations from monitors globally and regionally. We withheld 10% to 99% of monitors for testing to evaluate the sensitivity and robustness of model performance to the density of ground-based monitors. The model incorporating the geophysical a priori PM2.5 concentrations remains highly consistent with observations globally even under extreme conditions (e.g., 1% for training, R2 = 0.73), while the model without exhibits weaker performance (1% for training, R2 = 0.51).

PMID:38751607 | PMC:PMC11092969 | DOI:10.1021/acsestair.3c00054

Front Mol Biosci. 2024 May 1;11:1363384. doi: 10.3389/fmolb.2024.1363384. eCollection 2024.

ABSTRACT

BACKGROUND: Capillary ultrastructure in human skeletal muscles is dynamic and prone to alterations in response to many stimuli, e.g., systemic pathologies such as diabetes mellitus and arterial hypertension. Using transmission electron microscopy (TEM) images, several studies have been conducted to quantify the capillary ultrastructure by means of morphometry. Deep learning techniques like convolutional neural networks (CNNs) are utilized to extract data-driven characteristics and to recognize patterns. Hence, the aim of this study was to train a CNN to identify morphometric patterns that differ between capillaries in muscle biopsies of healthy participants and patients with systemic pathologies for the purpose of hypothesis generation.

METHODS: In this retrospective study we used 1810 electron micrographs from human skeletal muscle capillaries derived from 70 study participants which were classified as "healthy" controls or "patients" in dependence of the absence or presence of a documented history of diabetes mellitus, arterial hypertension or peripheral arterial disease. Using these micrographs, a pre-trained open-access CNN (ResNet101) was trained to discriminate between micrographs of capillaries of the two groups. The CNN with the highest diagnostic accuracies during training were subsequently compared with manual quantitative analysis of the capillary ultrastructure to distinguish between "healthy" controls and patients.

RESULTS: Using classification into controls or patients as allocation reference, receiver-operating-characteristics (ROC)-analysis of manually obtained BM thickness showed the best diagnostic accuracy of all morphometric indicators (area under the ROC-curve (AUC): 0.657 ± 0.050). The best performing CNN demonstrated a diagnostic accuracy of 79% (sensitivity 93%, specificity 92%). DeLong-Test of the ROC-curves showed a significant difference (p < 0.001) between the AUC of the best performing CNN and the BM thickness. The underlying morphology responsible for the network prediction focuses mainly on debridement of pericytes.

CONCLUSION: The hypothesis-generating approach using pretrained CNN distinguishes between capillaries depicted on electron micrographs of "healthy" controls and participants with a systemic pathology more accurately than by commonly used morphometric analysis.

PMID:38751446 | PMC:PMC11094256 | DOI:10.3389/fmolb.2024.1363384

BMC Med Inform Decis Mak. 2024 May 15;24(1):124. doi: 10.1186/s12911-024-02502-6.

ABSTRACT

BACKGROUND: Spatial molecular profiling depends on accurate cell segmentation. Identification and quantitation of individual cells in dense tissues, e.g. highly inflamed tissue caused by viral infection or immune reaction, remains a challenge.

METHODS: We first assess the performance of 18 deep learning-based cell segmentation models, either pre-trained or trained by us using two public image sets, on a set of immunofluorescence images stained with immune cell surface markers in skin tissue obtained during human herpes simplex virus (HSV) infection. We then further train eight of these models using up to 10,000+ training instances from the current image set. Finally, we seek to improve performance by tuning parameters of the most successful method from the previous step.

RESULTS: The best model before fine-tuning achieves a mean Average Precision (mAP) of 0.516. Prediction performance improves substantially after training. The best model is the cyto model from Cellpose. After training, it achieves an mAP of 0.694; with further parameter tuning, the mAP reaches 0.711.

CONCLUSION: Selecting the best model among the existing approaches and further training the model with images of interest produce the most gain in prediction performance. The performance of the resulting model compares favorably to human performance. The imperfection of the final model performance can be attributed to the moderate signal-to-noise ratio in the imageset.

PMID:38750526 | DOI:10.1186/s12911-024-02502-6

BMC Med Imaging. 2024 May 15;24(1):110. doi: 10.1186/s12880-024-01261-0.

ABSTRACT

Brain tumor classification using MRI images is a crucial yet challenging task in medical imaging. Accurate diagnosis is vital for effective treatment planning but is often hindered by the complex nature of tumor morphology and variations in imaging. Traditional methodologies primarily rely on manual interpretation of MRI images, supplemented by conventional machine learning techniques. These approaches often lack the robustness and scalability needed for precise and automated tumor classification. The major limitations include a high degree of manual intervention, potential for human error, limited ability to handle large datasets, and lack of generalizability to diverse tumor types and imaging conditions.To address these challenges, we propose a federated learning-based deep learning model that leverages the power of Convolutional Neural Networks (CNN) for automated and accurate brain tumor classification. This innovative approach not only emphasizes the use of a modified VGG16 architecture optimized for brain MRI images but also highlights the significance of federated learning and transfer learning in the medical imaging domain. Federated learning enables decentralized model training across multiple clients without compromising data privacy, addressing the critical need for confidentiality in medical data handling. This model architecture benefits from the transfer learning technique by utilizing a pre-trained CNN, which significantly enhances its ability to classify brain tumors accurately by leveraging knowledge gained from vast and diverse datasets.Our model is trained on a diverse dataset combining figshare, SARTAJ, and Br35H datasets, employing a federated learning approach for decentralized, privacy-preserving model training. The adoption of transfer learning further bolsters the model's performance, making it adept at handling the intricate variations in MRI images associated with different types of brain tumors. The model demonstrates high precision (0.99 for glioma, 0.95 for meningioma, 1.00 for no tumor, and 0.98 for pituitary), recall, and F1-scores in classification, outperforming existing methods. The overall accuracy stands at 98%, showcasing the model's efficacy in classifying various tumor types accurately, thus highlighting the transformative potential of federated learning and transfer learning in enhancing brain tumor classification using MRI images.

PMID:38750436 | DOI:10.1186/s12880-024-01261-0

J Imaging Inform Med. 2024 May 15. doi: 10.1007/s10278-024-01121-x. Online ahead of print.

ABSTRACT

OBJECTIVES: To preoperatively predict the high expression of Ki67 and positive pituitary transcription factor 1 (PIT-1) simultaneously in pituitary adenoma (PA) using three different radiomics models.

METHODS: A total of 247 patients with PA (training set: n = 198; test set: n = 49) were included in this retrospective study. The imaging features were extracted from preoperative contrast-enhanced T1WI (T1CE), T1-weighted imaging (T1WI), and T2-weighted imaging (T2WI). Feature selection was performed using Spearman's rank correlation coefficient and least absolute shrinkage and selection operator (LASSO). The classic machine learning (CML), deep learning (DL), and deep learning radiomics (DLR) models were constructed using logistic regression (LR), support vector machine (SVM), and multi-layer perceptron (MLP) algorithms. The area under the receiver operating characteristic (ROC) curve (AUC), sensitivity, specificity, accuracy, negative predictive value (NPV) and positive predictive value (PPV) were calculated for the training and test sets. In addition, combined with clinical characteristics, the best CML and the best DL models (SVM classifier), the DL radiomics nomogram (DLRN) was constructed to aid clinical decision-making.

RESULTS: Seven CML features, 96 DL features, and 107 DLR features were selected to construct CML, DL and DLR models. Compared to CML and DL model, the DLR model had the best performance. The AUC, sensitivity, specificity, accuracy, NPV and PPV were 0.827, 0.792, 0.800, 0.796, 0.800 and 0.792 in the test set, respectively.

CONCLUSIONS: Compared with CML and DL models, the DLR model shows the best performance in predicting the Ki67 and PIT-1 expression in PAs simultaneously.

PMID:38750186 | DOI:10.1007/s10278-024-01121-x