Sci Rep. 2024 Feb 1;14(1):2690. doi: 10.1038/s41598-024-53311-w.

ABSTRACT

Deep learning technology can effectively assist physicians in diagnosing chest radiographs. Conventional domain adaptation methods suffer from inaccurate lesion region localization, large errors in feature extraction, and a large number of model parameters. To address these problems, we propose a novel domain-adaptive method WDDM to achieve abnormal identification of chest radiographic images by combining Wasserstein distance and difference measures. Specifically, our method uses BiFormer as a multi-scale feature extractor to extract deep feature representations of data samples, which focuses more on discriminant features than convolutional neural networks and Swin Transformer. In addition, based on the loss minimization of Wasserstein distance and contrast domain differences, the source domain samples closest to the target domain are selected to achieve similarity and dissimilarity across domains. Experimental results show that compared with the non-transfer method that directly uses the network trained in the source domain to classify the target domain, our method has an average AUC increase of 14.8% and above. In short, our method achieves higher classification accuracy and better generalization performance.

PMID:38302556 | DOI:10.1038/s41598-024-53311-w

Sci Rep. 2024 Feb 1;14(1):2702. doi: 10.1038/s41598-024-53069-1.

ABSTRACT

In the healthcare sector, the health status and biological, and physical activity of the patient are monitored among different sensors that collect the required information about these activities using Wireless body area network (WBAN) architecture. Sensor-based human activity recognition (HAR), which offers remarkable qualities of ease and privacy, has drawn increasing attention from researchers with the growth of the Internet of Things (IoT) and wearable technology. Deep learning has the ability to extract high-dimensional information automatically, making end-to-end learning. The most significant obstacles to computer vision, particularly convolutional neural networks (CNNs), are the effect of the environment background, camera shielding, and other variables. This paper aims to propose and develop a new HAR system in WBAN dependence on the Gramian angular field (GAF) and DenseNet. Once the necessary signals are obtained, the input signals undergo pre-processing through artifact removal and median filtering. In the initial stage, the time series data captured by the sensors undergoes a conversion process, transforming it into 2-dimensional images by using the GAF algorithm. Then, DenseNet automatically makes the processes and integrates the data collected from diverse sensors. The experiment results show that the proposed method achieves the best outcomes in which it achieves 97.83% accuracy, 97.83% F-measure, and 97.64 Matthews correlation coefficient (MCC).

PMID:38302545 | DOI:10.1038/s41598-024-53069-1

Sci Rep. 2024 Feb 1;14(1):2633. doi: 10.1038/s41598-024-53107-y.

ABSTRACT

Heart diseases are leading to death across the globe. Exact detection and treatment for heart disease in its early stages could potentially save lives. Electrocardiogram (ECG) is one of the tests that take measures of heartbeat fluctuations. The deviation in the signals from the normal sinus rhythm and different variations can help detect various heart conditions. This paper presents a novel approach to cardiac disease detection using an automated Convolutional Neural Network (CNN) system. Leveraging the Scale-Invariant Feature Transform (SIFT) for unique ECG signal image feature extraction, our model classifies signals into three categories: Arrhythmia (ARR), Congestive Heart Failure (CHF), and Normal Sinus Rhythm (NSR). The proposed model has been evaluated using 96 Arrhythmia, 30 CHF, and 36 NSR ECG signals, resulting in a total of 162 images for classification. Our proposed model achieved 99.78% accuracy and an F1 score of 99.78%, which is among one of the highest in the models which were recorded to date with this dataset. Along with the SIFT, we also used HOG and SURF techniques individually and applied the CNN model which achieved 99.45% and 78% accuracy respectively which proved that the SIFT-CNN model is a well-trained and performed model. Notably, our approach introduces significant novelty by combining SIFT with a custom CNN model, enhancing classification accuracy and offering a fresh perspective on cardiac arrhythmia detection. This SIFT-CNN model performed exceptionally well and better than all existing models which are used to classify heart diseases.

PMID:38302520 | DOI:10.1038/s41598-024-53107-y

Nat Commun. 2024 Feb 1;15(1):976. doi: 10.1038/s41467-024-44930-y.

ABSTRACT

Early detection is critical to achieving improved treatment outcomes for child patients with congenital heart diseases (CHDs). Therefore, developing effective CHD detection techniques using low-cost and non-invasive pediatric electrocardiogram are highly desirable. We propose a deep learning approach for CHD detection, CHDdECG, which automatically extracts features from pediatric electrocardiogram and wavelet transformation characteristics, and integrates them with key human-concept features. Developed on 65,869 cases, CHDdECG achieved ROC-AUC of 0.915 and specificity of 0.881 on a real-world test set covering 12,000 cases. Additionally, on two external test sets with 7137 and 8121 cases, the overall ROC-AUC were 0.917 and 0.907 while specificities were 0.937 and 0.907. Notably, CHDdECG surpassed cardiologists in CHD detection performance comparison, and feature importance scores suggested greater influence of automatically extracted electrocardiogram features on CHD detection compared with human-concept features, implying that CHDdECG may grasp some knowledge beyond human cognition. Our study directly impacts CHD detection with pediatric electrocardiogram and demonstrates the potential of pediatric electrocardiogram for broader benefits.

PMID:38302502 | DOI:10.1038/s41467-024-44930-y

Sci Data. 2024 Feb 1;11(1):155. doi: 10.1038/s41597-024-02975-0.

ABSTRACT

Urinary tract infection (UTI) is a common disorder. Its diagnosis can be made by microscopic examination of voided urine for markers of infection. This manual technique is technically difficult, time-consuming and prone to inter-observer errors. The application of computer vision to this domain has been slow due to the lack of a clinical image dataset from UTI patients. We present an open dataset containing 300 images and 3,562 manually annotated urinary cells labelled into seven classes of clinically significant cell types. It is an enriched dataset acquired from the unstained and untreated urine of patients with symptomatic UTI using a simple imaging system. We demonstrate that this dataset can be used to train a Patch U-Net, a novel deep learning architecture with a random patch generator to recognise urinary cells. Our hope is, with this dataset, UTI diagnosis will be made possible in nearly all clinical settings by using a simple imaging system which leverages advanced machine learning techniques.

PMID:38302487 | DOI:10.1038/s41597-024-02975-0

J Chem Inf Model. 2024 Feb 1. doi: 10.1021/acs.jcim.3c01964. Online ahead of print.

ABSTRACT

Deep learning-based de novo molecular design has recently gained significant attention. While numerous DL-based generative models have been successfully developed for designing novel compounds, the majority of the generated molecules lack sufficiently novel scaffolds or high drug-like profiles. The aforementioned issues may not be fully captured by commonly used metrics for the assessment of molecular generative models, such as novelty, diversity, and quantitative estimation of the drug-likeness score. To address these limitations, we proposed a genetic algorithm-guided generative model called GARel (genetic algorithm-based receptor-ligand interaction generator), a novel framework for training a DL-based generative model to produce drug-like molecules with novel scaffolds. To efficiently train the GARel model, we utilized dense net to update the parameters based on molecules with novel scaffolds and drug-like features. To demonstrate the capability of the GARel model, we used it to design inhibitors for three targets: AA2AR, EGFR, and SARS-Cov2. The results indicate that GARel-generated molecules feature more diverse and novel scaffolds and possess more desirable physicochemical properties and favorable docking scores. Compared with other generative models, GARel makes significant progress in balancing novelty and drug-likeness, providing a promising direction for the further development of DL-based de novo design methodology with potential impacts on drug discovery.

PMID:38302422 | DOI:10.1021/acs.jcim.3c01964

J Neurointerv Surg. 2024 Feb 1:jnis-2023-021283. doi: 10.1136/jnis-2023-021283. Online ahead of print.

ABSTRACT

BACKGROUND: Outlining acutely infarcted tissue on non-contrast CT is a challenging task for which human inter-reader agreement is limited. We explored two different methods for training a supervised deep learning algorithm: one that used a segmentation defined by majority vote among experts and another that trained randomly on separate individual expert segmentations.

METHODS: The data set consisted of 260 non-contrast CT studies in 233 patients with acute ischemic stroke recruited from the multicenter DEFUSE 3 (Endovascular Therapy Following Imaging Evaluation for Ischemic Stroke 3) trial. Additional external validation was performed using 33 patients with matched stroke onset times from the University Hospital Lausanne. A benchmark U-Net was trained on the reference annotations of three experienced neuroradiologists to segment ischemic brain tissue using majority vote and random expert sampling training schemes. The median of volume, overlap, and distance segmentation metrics were determined for agreement in lesion segmentations between (1) three experts, (2) the majority model and each expert, and (3) the random model and each expert. The two sided Wilcoxon signed rank test was used to compare performances (1) to 2) and (1) to (3). We further compared volumes with the 24 hour follow-up diffusion weighted imaging (DWI, final infarct core) and correlations with clinical outcome (modified Rankin Scale (mRS) at 90 days) with the Spearman method.

RESULTS: The random model outperformed the inter-expert agreement ((1) to (2)) and the majority model ((1) to (3)) (dice 0.51±0.04 vs 0.36±0.05 (P<0.0001) vs 0.45±0.05 (P<0.0001)). The random model predicted volume correlated with clinical outcome (0.19, P<0.05), whereas the median expert volume and majority model volume did not. There was no significant difference when comparing the volume correlations between random model, median expert volume, and majority model to 24 hour follow-up DWI volume (P>0.05, n=51).

CONCLUSION: The random model for ischemic injury delineation on non-contrast CT surpassed the inter-expert agreement ((1) to (2)) and the performance of the majority model ((1) to (3)). We showed that the random model volumetric measures of the model were consistent with 24 hour follow-up DWI.

PMID:38302420 | DOI:10.1136/jnis-2023-021283

J Chem Inf Model. 2024 Feb 1. doi: 10.1021/acs.jcim.3c01923. Online ahead of print.

ABSTRACT

We present a quantum mechanical/machine learning (ML) framework based on random forest to accurately predict the pKas of complex organic molecules using inexpensive density functional theory (DFT) calculations. By including physics-based features from low-level DFT calculations and structural features from our connectivity-based hierarchy (CBH) fragmentation protocol, we can correct the systematic error associated with DFT. The generalizability and performance of our model are evaluated on two benchmark sets (SAMPL6 and Novartis). We believe the carefully curated input of physics-based features lessens the model's data dependence and need for complex deep learning architectures, without compromising the accuracy of the test sets. As a point of novelty, our work extends the applicability of CBH, employing it for the generation of viable molecular descriptors for ML.

PMID:38301279 | DOI:10.1021/acs.jcim.3c01923

Neurosurg Focus. 2024 Feb;56(2):E5. doi: 10.3171/2023.11.FOCUS23685.

ABSTRACT

OBJECTIVE: Contemporary oncological paradigms for adjuvant treatment of low- and intermediate-grade gliomas are often guided by a limited array of parameters, overlooking the dynamic nature of the disease. The authors' aim was to develop a comprehensive multivariate glioma growth model based on multicentric data, to facilitate more individualized therapeutic strategies.

METHODS: Random slope models with subject-specific random intercepts were fitted to a retrospective cohort of grade II and III gliomas from the database at Kepler University Hospital (n = 191) to predict future mean tumor diameters. Deep learning-based radiomics was used together with a comprehensive clinical dataset and evaluated on an external prospectively collected validation cohort from University Hospital Zurich (n = 9). Prediction quality was assessed via mean squared prediction error.

RESULTS: A mean squared prediction error of 0.58 cm for the external validation cohort was achieved, indicating very good prognostic value. The mean ± SD time to adjuvant therapy was 28.7 ± 43.3 months and 16.1 ± 14.6 months for the training and validation cohort, respectively, with a mean of 6.2 ± 5 and 3.6 ± 0.7, respectively, for number of observations. The observed mean tumor diameter per year was 0.38 cm (95% CI 0.25-0.51) for the training cohort, and 1.02 cm (95% CI 0.78-2.82) for the validation cohort. Glioma of the superior frontal gyrus showed a higher rate of tumor growth than insular glioma. Oligodendroglioma showed less pronounced growth, anaplastic astrocytoma-unlike anaplastic oligodendroglioma-was associated with faster tumor growth. Unlike the impact of extent of resection, isocitrate dehydrogenase (IDH) had negligible influence on tumor growth. Inclusion of radiomics variables significantly enhanced the prediction performance of the random slope model used.

CONCLUSIONS: The authors developed an advanced statistical model to predict tumor volumes both pre- and postoperatively, using comprehensive data prior to the initiation of adjuvant therapy. Using radiomics enhanced the precision of the prediction models. Whereas tumor extent of resection and topology emerged as influential factors in tumor growth, the IDH status did not. This study emphasizes the imperative of advanced computational methods in refining personalized low-grade glioma treatment, advocating a move beyond traditional paradigms.

PMID:38301234 | DOI:10.3171/2023.11.FOCUS23685

J Chem Inf Model. 2024 Feb 1. doi: 10.1021/acs.jcim.3c01471. Online ahead of print.

ABSTRACT

Protein-molecule interactions play a crucial role in various biological functions, with their accurate prediction being pivotal for drug discovery and design processes. Traditional methods for predicting protein-molecule interactions are limited. Some can only predict interactions with a specific molecule, restricting their applicability, while others target multiple molecule types but fail to efficiently process diverse interaction information, leading to complexity and inefficiency. This study presents a novel deep learning model, MucLiPred, equipped with a dual contrastive learning mechanism aimed at improving the prediction of multiple molecule-protein interactions and the identification of potential molecule-binding residues. The residue-level paradigm focuses on differentiating binding from non-binding residues, illuminating detailed local interactions. The type-level paradigm, meanwhile, analyzes overarching contexts of molecule types, like DNA or RNA, ensuring that representations of identical molecule types gravitate closer in the representational space, bolstering the model's proficiency in discerning interaction motifs. This dual approach enables comprehensive multi-molecule predictions, elucidating the relationships among different molecule types and strengthening precise protein-molecule interaction predictions. Empirical evidence demonstrates MucLiPred's superiority over existing models in robustness and prediction accuracy. The integration of dual contrastive learning techniques amplifies its capability to detect potential molecule-binding residues with precision. Further optimization, separating representational and classification tasks, has markedly improved its performance. MucLiPred thus represents a significant advancement in protein-molecule interaction prediction, setting a new precedent for future research in this field.

PMID:38301174 | DOI:10.1021/acs.jcim.3c01471

Proc Natl Acad Sci U S A. 2024 Feb 6;121(6):e2300838121. doi: 10.1073/pnas.2300838121. Epub 2024 Feb 1.

ABSTRACT

Proteins play a central role in biology from immune recognition to brain activity. While major advances in machine learning have improved our ability to predict protein structure from sequence, determining protein function from its sequence or structure remains a major challenge. Here, we introduce holographic convolutional neural network (H-CNN) for proteins, which is a physically motivated machine learning approach to model amino acid preferences in protein structures. H-CNN reflects physical interactions in a protein structure and recapitulates the functional information stored in evolutionary data. H-CNN accurately predicts the impact of mutations on protein stability and binding of protein complexes. Our interpretable computational model for protein structure-function maps could guide design of novel proteins with desired function.

PMID:38300863 | DOI:10.1073/pnas.2300838121

IEEE Trans Pattern Anal Mach Intell. 2024 Feb 1;PP. doi: 10.1109/TPAMI.2024.3361155. Online ahead of print.

ABSTRACT

One fundamental problem in deep learning is understanding the excellent performance of deep Neural Networks (NNs) in practice. An explanation for the superiority of NNs is that they can realize a large family of complicated functions, i.e., they have powerful expressivity. The expressivity of a Neural Network with Piecewise Linear activations (PLNN) can be quantified by the maximal number of linear regions it can separate its input space into. In this paper, we provide several mathematical results needed for studying the linear regions of Convolutional Neural Networks with Piecewise Linear activations (PLCNNs), and use them to derive the maximal and average numbers of linear regions for one-layer PLCNNs. Furthermore, we obtain upper and lower bounds for the number of linear regions of multi-layer PLCNNs. Our results suggest that deeper PLCNNs have more powerful expressivity than shallow PLCNNs, while PLCNNs have more expressivity than fully-connected PLNNs per parameter, in terms of the number of linear regions.

PMID:38300783 | DOI:10.1109/TPAMI.2024.3361155

IEEE Trans Image Process. 2024 Feb 1;PP. doi: 10.1109/TIP.2024.3359066. Online ahead of print.

ABSTRACT

Hashing and quantization have greatly succeeded by benefiting from deep learning for large-scale image retrieval. Recently, deep product quantization methods have attracted wide attention. However, representation capability of codewords needs to be further improved. Moreover, since the number of codewords in the codebook depends on experience, representation capability of codewords is usually imbalanced, which leads to redundancy or insufficiency of codewords and reduces retrieval performance. Therefore, in this paper, we propose a novel deep product quantization method, named Entropy Optimized deep Weighted Product Quantization (EOWPQ), which not only encodes samples into the weighted codewords in a new flexible manner but also balances the codeword assignment, improving while balancing representation capability of codewords. Specifically, we encode samples using the linear weighted sum of codewords instead of a single codeword as traditionally. Meanwhile, we establish the linear relationship between the weighted codewords and semantic labels, which effectively maintains semantic information of codewords. Moreover, in order to balance the codeword assignment, that is, avoiding some codewords representing most samples or some codewords representing very few samples, we maximize the entropy of the coding probability distribution and obtain the optimal coding probability distribution of samples by utilizing optimal transport theory, which achieves the optimal assignment of codewords and balances representation capability of codewords. The experimental results on three benchmark datasets show that EOWPQ can achieve better retrieval performance and also show the improvement of representation capability of codewords and the balance of codeword assignment.

PMID:38300776 | DOI:10.1109/TIP.2024.3359066

Transl Vis Sci Technol. 2024 Feb 1;13(2):1. doi: 10.1167/tvst.13.2.1.

ABSTRACT

PURPOSE: Artificial intelligence (AI)-assisted ultra-widefield (UWF) fundus photographic interpretation is beneficial to improve the screening of fundus abnormalities. Therefore we constructed an AI machine-learning approach and performed preliminary training and validation.

METHODS: We proposed a two-stage deep learning-based framework to detect early retinal peripheral degeneration using UWF images from the Chinese Air Force cadets' medical selection between February 2016 and June 2022. We developed a detection model for the localization of optic disc and macula, which are used to find the peripheral areas. Then we developed six classification models for the screening of various retinal cases. We also compared our proposed framework with two baseline models reported in the literature. The performance of the screening models was evaluated by area under the receiver operating curve (AUC) with 95% confidence interval.

RESULTS: A total of 3911 UWF fundus images were used to develop the deep learning model. The external validation included 760 UWF fundus images. The results of comparison study revealed that our proposed framework achieved competitive performance compared to existing baselines while also demonstrating significantly faster inference time. The developed classification models achieved an average AUC of 0.879 on six different retinal cases in the external validation dataset.

CONCLUSIONS: Our two-stage deep learning-based framework improved the machine learning efficiency of the AI model for fundus images with high resolution and many interference factors by maximizing the retention of valid information and compressing the image file size.

TRANSLATIONAL RELEVANCE: This machine learning model may become a new paradigm for developing UWF fundus photography AI-assisted diagnosis.

PMID:38300623 | DOI:10.1167/tvst.13.2.1

Brief Bioinform. 2024 Jan 22;25(2):bbad523. doi: 10.1093/bib/bbad523.

ABSTRACT

Accurate cell type annotation in single-cell RNA-sequencing data is essential for advancing biological and medical research, particularly in understanding disease progression and tumor microenvironments. However, existing methods are constrained by single feature extraction approaches, lack of adaptability to immune cell types with similar molecular profiles but distinct functions and a failure to account for the impact of cell label noise on model accuracy, all of which compromise the precision of annotation. To address these challenges, we developed a supervised approach called scMMT. We proposed a novel feature extraction technique to uncover more valuable information. Additionally, we constructed a multi-task learning framework based on the GradNorm method to enhance the recognition of challenging immune cells and reduce the impact of label noise by facilitating mutual reinforcement between cell type annotation and protein prediction tasks. Furthermore, we introduced logarithmic weighting and label smoothing mechanisms to enhance the recognition ability of rare cell types and prevent model overconfidence. Through comprehensive evaluations on multiple public datasets, scMMT has demonstrated state-of-the-art performance in various aspects including cell type annotation, rare cell identification, dropout and label noise resistance, protein expression prediction and low-dimensional embedding representation.

PMID:38300515 | DOI:10.1093/bib/bbad523

J Neurooncol. 2024 Feb 1. doi: 10.1007/s11060-024-04580-y. Online ahead of print.

ABSTRACT

PURPOSE: Close MRI surveillance of patients with brain metastases following Stereotactic Radiosurgery (SRS) treatment is essential for assessing treatment response and the current disease status in the brain. This follow-up necessitates the comparison of target lesion sizes in pre- (prior) and post-SRS treatment (current) T1W-Gad MRI scans. Our aim was to evaluate SimU-Net, a novel deep-learning model for the detection and volumetric analysis of brain metastases and their temporal changes in paired prior and current scans.

METHODS: SimU-Net is a simultaneous multi-channel 3D U-Net model trained on pairs of registered prior and current scans of a patient. We evaluated its performance on 271 pairs of T1W-Gad MRI scans from 226 patients who underwent SRS. An expert oncological neurosurgeon manually delineated 1,889 brain metastases in all the MRI scans (1,368 with diameters > 5 mm, 834 > 10 mm). The SimU-Net model was trained/validated on 205 pairs from 169 patients (1,360 metastases) and tested on 66 pairs from 57 patients (529 metastases). The results were then compared to the ground truth delineations.

RESULTS: SimU-Net yielded a mean (std) detection precision and recall of 1.00±0.00 and 0.99±0.06 for metastases > 10 mm, 0.90±0.22 and 0.97±0.12 for metastases > 5 mm of, and 0.76±0.27 and 0.94±0.16 for metastases of all sizes. It improves lesion detection precision by 8% for all metastases sizes and by 12.5% for metastases < 10 mm with respect to standalone 3D U-Net. The segmentation Dice scores were 0.90±0.10, 0.89±0.10 and 0.89±0.10 for the above metastases sizes, all above the observer variability of 0.80±0.13.

CONCLUSION: Automated detection and volumetric quantification of brain metastases following SRS have the potential to enhance the assessment of treatment response and alleviate the clinician workload.

PMID:38300389 | DOI:10.1007/s11060-024-04580-y

Eur Radiol. 2024 Feb 1. doi: 10.1007/s00330-023-10452-2. Online ahead of print.

ABSTRACT

OBJECTIVES: This study aims to develop computer-aided detection (CAD) for colorectal cancer (CRC) using abdominal CT based on a deep convolutional neural network.

METHODS: This retrospective study included consecutive patients with colorectal adenocarcinoma who underwent abdominal CT before CRC resection surgery (training set = 379, test set = 103). We customized the 3D U-Net of nnU-Net (CUNET) for CRC detection, which was trained with fivefold cross-validation using annotated CT images. CUNET was validated using datasets covering various clinical situations and institutions: an internal test set (n = 103), internal patients with CRC first determined by CT (n = 54) and asymptomatic CRC (n = 51), and an external validation set from two institutions (n = 60). During each validation, data from the healthy population were added (internal = 60; external = 130). CUNET was compared with other deep CNNs: residual U-Net and EfficientDet. The CAD performances were evaluated using per-CRC sensitivity (true positive/all CRCs), free-response receiver operating characteristic (FROC), and jackknife alternative FROC (JAFROC) curves.

RESULTS: CUNET showed a higher maximum per-CRC sensitivity than residual U-Net and EfficientDet (internal test set 91.3% vs. 61.2%, and 64.1%). The per-CRC sensitivity of CUNET at false-positive rates of 3.0 was as follows: internal CRC determined by CT, 89.3%; internal asymptomatic CRC, 87.3%; and external validation, 89.6%. CUNET detected 69.2% (9/13) of CRCs missed by radiologists and 89.7% (252/281) of CRCs from all validation sets.

CONCLUSIONS: CUNET can detect CRC on abdominal CT in patients with various clinical situations and from external institutions.

KEY POINTS: • Customized 3D U-Net of nnU-Net (CUNET) can be applied to the opportunistic detection of colorectal cancer (CRC) in abdominal CT, helping radiologists detect unexpected CRC. • CUNET showed the best performance at false-positive rates ≥ 3.0, and 30.1% of false-positives were in the colorectum. CUNET detected 69.2% (9/13) of CRCs missed by radiologists and 87.3% (48/55) of asymptomatic CRCs. • CUNET detected CRCs in multiple validation sets composed of varying clinical situations and from different institutions, and CUNET detected 89.7% (252/281) of CRCs from all validation sets.

PMID:38300293 | DOI:10.1007/s00330-023-10452-2

Cornea. 2024 Feb 1. doi: 10.1097/ICO.0000000000003467. Online ahead of print.

ABSTRACT

PURPOSE: The prevalence of keratoconus in the general population is reported to be up to 1 of 84. Over the past 2 decades, diagnosis and management evolved rapidly, but keratoconus screening in clinical practice is still challenging and asks for improving the accuracy of keratoconus detection. Deep learning (DL) offers considerable promise for improving the accuracy and speed of medical imaging interpretation. We establish an inventory of studies conducted with DL algorithms that have attempted to diagnose keratoconus.

METHODS: This systematic review was conducted according to the recommendations of the PRISMA statement. We searched (Pre-)MEDLINE, Embase, Science Citation Index, Conference Proceedings Citation Index, arXiv document server, and Google Scholar from inception to February 18, 2022. We included studies that evaluated the performance of DL algorithms in the diagnosis of keratoconus. The main outcome was diagnostic performance measured as sensitivity and specificity, and the methodological quality of the included studies was assessed using QUADAS-2.

RESULTS: Searches retrieved 4100 nonduplicate records, and we included 19 studies in the qualitative synthesis and 10 studies in the exploratory meta-analysis. The overall study quality was limited because of poor reporting of patient selection and the use of inadequate reference standards. We found a pooled sensitivity of 97.5% (95% confidence interval, 93.6%-99.0%) and a pooled specificity of 97.2% (95% confidence interval, 85.7%-99.5%) for topography images as input.

CONCLUSIONS: Our systematic review found that the overall diagnostic performance of DL models to detect keratoconus was good, but the methodological quality of included studies was modest.

PMID:38300179 | DOI:10.1097/ICO.0000000000003467

J Synchrotron Radiat. 2024 Mar 1. doi: 10.1107/S1600577524000146. Online ahead of print.

ABSTRACT

The structural and chemical evolution of battery electrodes at the nanoscale plays an important role in affecting the cell performance. Nano-resolution X-ray microscopy has been demonstrated as a powerful technique for characterizing the evolution of battery electrodes under operating conditions with sensitivity to their morphology, compositional distribution and redox heterogeneity. In real-world batteries, the electrode could deform upon battery operation, causing challenges for the image registration which is necessary for several experimental modalities, e.g. XANES imaging. To address this challenge, this work develops a deep-learning-based method for automatic particle identification and tracking. This approach was not only able to facilitate image registration with good robustness but also allowed quantification of the degree of sample deformation. The effectiveness of the method was first demonstrated using synthetic datasets with known ground truth. The method was then applied to an experimental dataset collected on an operating lithium battery cell, revealing a high degree of intra- and interparticle chemical complexity in operating batteries.

PMID:38300132 | DOI:10.1107/S1600577524000146

Opt Lett. 2024 Feb 1;49(3):602-605. doi: 10.1364/OL.506775.

ABSTRACT

High dynamic range (HDR) 3D measurement is a meaningful but challenging problem. Recently, many deep-learning-based methods have been proposed for the HDR problem. However, due to learning redundant fringe intensity information, their networks are difficult to converge for data with complex surface reflectivity and various illumination conditions, resulting in non-robust performance. To address this problem, we propose a physics-based supervised learning method. By introducing the physical model for phase retrieval, we design a novel, to the best of our knowledge, sinusoidal-component-to-sinusoidal-component mapping paradigm. Consequently, the scale difference of fringe intensity in various illumination scenarios can be eliminated. Compared with conventional supervised-learning methods, our method can greatly promote the convergence of the network and the generalization ability, while compared with the recently proposed unsupervised-learning method, our method can recover complex surfaces with much more details. To better evaluate our method, we specially design the experiment by training the network merely using the metal objects and testing the performance using different diffuse sculptures, metal surfaces, and their hybrid scenes. Experiments for all the testing scenarios have high-quality phase recovery with an STD error of about 0.03 rad, which reveals the superior generalization ability for complex reflectivity and various illumination conditions. Furthermore, the zoom-in 3D plots of the sculpture verify its fidelity on recovering fine details.

PMID:38300069 | DOI:10.1364/OL.506775