Efficacy of MRI data harmonization in the age of machine learning: a multicenter study across 36 datasets

Scientific Data, Journal Year: 2024, Volume and Issue: 11(1)

Published: Jan. 23, 2024

Abstract Pooling publicly-available MRI data from multiple sites allows to assemble extensive groups of subjects, increase statistical power, and promote reuse with machine learning techniques. The harmonization multicenter is necessary reduce the confounding effect associated non-biological sources variability in data. However, when applied entire dataset before learning, leads leakage, because information outside training set may affect model building, potentially falsely overestimate performance. We propose a 1) measurement efficacy harmonization; 2) harmonizer transformer, i.e., an implementation ComBat allowing its encapsulation among preprocessing steps pipeline, avoiding leakage by design. tested these tools using brain T 1 -weighted 1740 healthy subjects acquired at 36 sites. After harmonization, site was removed or reduced, we showed predicting individual age data, highlighting that introducing transformer into pipeline for

Language: Английский

---

Automated Detection of Keratorefractive Laser Surgeries on Optical Coherence Tomography Using Deep Learning

Journal of Refractive Surgery, Journal Year: 2025, Volume and Issue: 41(3)

Published: March 1, 2025

Purpose To report a deep learning neural network on anterior segment optical coherence tomography (AS-OCT) for automated detection of different keratorefractive laser surgeries—including in situ keratomileusis with femtosecond microkeratome (femto-LASIK), LASIK mechanical microkeratome, photorefractive keratectomy (PRK), lenticule extraction (KLEx), and non-operated eyes—while also distinguishing between myopic hyperopic treatments within these procedures. Methods A total 14,948 eye scans from 2,278 eyes 1,166 patients were used to develop algorithm an 80/10/10 patient distribution training, validation, testing phases, respectively. The was evaluated its accuracy, F1 scores, area under precision-recall curve (AUPRC), receiver operating characteristic (AUROC). Results On the test dataset, able detect surgical classes accuracy 96%, weighted-average score macro-average 96%. further subclasses each class, 90%, 83%. Conclusions Neural networks can accurately classify patient's history AS-OCT scans, which may support treatment planning, intraocular lens calculations, ectasia assessment, particularly cases where electronic health records are incomplete. This represents step toward transforming OCT diagnostic more comprehensive screening tool refractive clinics. [ J Refract Surg . 2025;41(3):e248–e256.]

Language: Английский

---

Predicting Adolescent Mental Health Outcomes Across Cultures: A Machine Learning Approach

Journal of Youth and Adolescence, Journal Year: 2023, Volume and Issue: 52(8), P. 1595 - 1619

Published: April 19, 2023

Language: Английский

---

Barometers Behaving Badly II: a Critical Evaluation of Cpx-Only and Cpx-Liq Thermobarometry in Variably-Hydrous Arc Magmas

Journal of Petrology, Journal Year: 2023, Volume and Issue: 64(8)

Published: July 5, 2023

ABSTRACT The chemistry of erupted clinopyroxene crystals (±equilibrium liquids) have been widely used to deduce the pressures and temperatures magma storage in volcanic arcs. However, large number different equations parameterizing relationship between mineral melt compositions intensive variables such as pressure temperature yield vastly results, with implications for our interpretation conditions. We use a new test dataset composed average Clinopyroxene-Liquid (Cpx-Liq) from N = 543 variably hydrous experiments at crustal conditions (1 bar 17 kbar) assess performance thermobarometers identify most accurate precise expressions application subduction zone magmas. First, we equilibrium tests, finding that comparing measured predicted Enstatite-Ferrosillite KD (using Fet both phases) are useful tests arc magmas, whereas CaTs, CaTi Jd limited utility. then apply further quality filters based on cation sums (3.95–4.05), analyses (N > 5) presence reported H2O data quenched experimental glass (hereafter ‘liquid’) obtain filtered 214). this compare calculated versus combinations thermobarometers. A Cpx-Liq thermometers perform very well when liquid contents known, although Cpx composition contributes little relative composition. Most Cpx-only badly, greatly overestimating experiments. These two findings demonstrate alone holds information systems. barometers show similar one another (mostly yielding root mean square errors [RMSEs] 2–3.5 kbar), best currently outperform barometers. also sensitivity contents, which poorly constrained many natural Overall, work demonstrates Cpx-based barometry individual only provides sufficient resolution distinguish broad regions continental arcs (e.g. upper, mid, lower crust). Significant averaging can reduce RMSEs ~1.3–1.9 kbar. hope motivate substantial amount analytical is required estimates depths ± Liq

Language: Английский

---

Deep Learning Approaches to Osteosarcoma Diagnosis and Classification: A Comparative Methodological Approach

Cancers, Journal Year: 2023, Volume and Issue: 15(8), P. 2290 - 2290

Published: April 13, 2023

Background: Osteosarcoma is the most common primary malignancy of bone, being prevalent in childhood and adolescence. Despite recent progress diagnostic methods, histopathology remains gold standard for disease staging therapy decisions. Machine learning deep methods have shown potential evaluating classifying histopathological cross-sections. Methods: This study used publicly available images osteosarcoma cross-sections to analyze compare performance state-of-the-art neural networks evaluation osteosarcomas. Results: The classification did not necessarily improve when using larger on our dataset. In fact, smallest network combined with image input size achieved best overall performance. When trained 5-fold cross-validation, MobileNetV2 91% accuracy. Conclusions: present highlights importance careful selection size. Our results indicate that a number parameters always better, can be smaller more efficient networks. identification an optimal training configuration could greatly accuracy diagnoses ultimately lead better outcomes patients.

Language: Английский

---

Super-resolution mapping of anisotropic tissue structure with diffusion MRI and deep learning

Scientific Reports, Journal Year: 2025, Volume and Issue: 15(1)

Published: Feb. 24, 2025

Abstract Diffusion magnetic resonance imaging (diffusion MRI) is widely employed to probe the diffusive motion of water molecules within tissue. Numerous diseases and processes affecting central nervous system can be detected monitored via diffusion MRI thanks its sensitivity microstructural alterations in The latter has prompted interest quantitative mapping parameters, such as fiber orientation distribution function (fODF), which instrumental for noninvasively underlying axonal tracts white matter through a procedure known tractography. However, applications demand repeated acquisitions volumes with varied experimental parameters demanding long acquisition times and/or limited spatial resolution. In this work, we present deep-learning-based approach increasing resolution data form fODFs obtained constrained spherical deconvolution. proposed evaluated on high quality from Human Connectome Project, shown generate upsampled results greater correspondence ground truth high-resolution than achieved ordinary spline interpolation methods. Furthermore, employ measure based earth mover’s distance assess accuracy fODFs. At low signal-to-noise ratios, our super-resolution method provides more accurate estimates fODF compared collected 8 smaller voxel volume.

Language: Английский

---

AutoPrognosis 2.0: Democratizing diagnostic and prognostic modeling in healthcare with automated machine learning

PLOS Digital Health, Journal Year: 2023, Volume and Issue: 2(6), P. e0000276 - e0000276

Published: June 22, 2023

Diagnostic and prognostic models are increasingly important in medicine inform many clinical decisions. Recently, machine learning approaches have shown improvement over conventional modeling techniques by better capturing complex interactions between patient covariates a data-driven manner. However, the use of introduces technical practical challenges that thus far restricted widespread adoption such settings. To address these empower healthcare professionals, we present an open-source framework, AutoPrognosis 2.0, to facilitate development diagnostic models. leverages state-of-the-art advances automated develop optimized pipelines, incorporates model explainability tools, enables deployment demonstrators, without requiring significant expertise. demonstrate provide illustrative application where construct risk score for diabetes using UK Biobank, prospective study 502,467 individuals. The produced our framework achieve greater discrimination than expert scores. We implemented as web-based decision support tool, which can be publicly accessed patients clinicians. By open-sourcing tool community, aim clinicians other medical practitioners with accessible resource new scores, personalized diagnostics, prognostics techniques. Software: https://github.com/vanderschaarlab/AutoPrognosis.

Language: Английский

---

ConvAttenMixer: Brain tumor detection and type classification using convolutional mixer with external and self-attention mechanisms

Journal of King Saud University - Computer and Information Sciences, Journal Year: 2023, Volume and Issue: 35(10), P. 101810 - 101810

Published: Oct. 21, 2023

Attention-based methods have recently demonstrated notable advancements in brain tumor classification. To further advance and strengthen this development, we developed ConvAttenMixer, a transformer model that incorporates convolutional layers along with two attention mechanisms: self-attention external attention. The proposed utilizes blocks of convolution mixers to effectively process blend across patches, thereby enhancing the model's ability capture spatial channel-wise dependencies MRI images. block enables prioritize important regions within image establish by assigning weights each part based on their relevance task. This allows emphasize crucial local features, disregard irrelevant ones, interactions between different patches. On other hand, focuses more significant global features captures among images, enabling correlations all samples. classification head is simple yet effective designed output feature maps using squeeze-and-excitation mechanism, which turn assigns higher channels suppresses less-relevant channels. For experimentation, our ConvAttenMixer was trained dataset consisting 5712 scans subsequently tested 1311 for into glioma, meningioma, pituitary tumor, no-tumor Different variants were evaluated. optimally performing architecture evaluated against state-of-the-art baselines, namely MLP, attention-based pooling net, mixer net. Extensive experiments outperformed employed either or mechanisms, while requiring significantly less computational memory. suggested exhibited precision, recall, f-measure, achieving highest accuracy 0.9794 compared baselines' accuracy, ranged from 0.87 0.93. demonstrates operate locally patch level globally sample attention, as well information channel mechanism.

Language: Английский

---

An investigation of machine learning algorithms and data augmentation techniques for diabetes diagnosis using class imbalanced BRFSS dataset

Healthcare Analytics, Journal Year: 2023, Volume and Issue: 5, P. 100297 - 100297

Published: Dec. 30, 2023

Diabetes is a prevalent chronic condition that poses significant challenges to early diagnosis and identifying at-risk individuals. Machine learning plays crucial role in diabetes detection by leveraging its ability process large volumes of data identify complex patterns. However, imbalanced data, where the number diabetic cases substantially smaller than non-diabetic cases, complicates identification individuals with using machine algorithms. This study focuses on predicting whether person at risk diabetes, considering individual's health socio-economic conditions while mitigating posed data. We employ several augmentation techniques, such as oversampling (Synthetic Minority Over Sampling for Nominal Data, i.e.SMOTE-N), undersampling (Edited Nearest Neighbor, i.e. ENN), hybrid sampling techniques (SMOTE-Tomek SMOTE-ENN) training before applying algorithms minimize impact Our sheds light significance carefully utilizing without any leakage enhance effectiveness Moreover, it offers complete structure healthcare practitioners, from obtaining prediction, enabling them make informed decisions.

Language: Английский

---

AI chatbots show promise but limitations on UK medical exam questions: a comparative performance study

Scientific Reports, Journal Year: 2024, Volume and Issue: 14(1)

Published: Aug. 14, 2024

Abstract Large language models (LLMs) like ChatGPT have potential applications in medical education such as helping students study for their licensing exams by discussing unclear questions with them. However, they require evaluation on these complex tasks. The purpose of this was to evaluate how well publicly accessible LLMs performed simulated UK board exam questions. 423 board-style from 9 (MRCS, MRCP, etc.) were answered seven (ChatGPT-3.5, ChatGPT-4, Bard, Perplexity, Claude, Bing, Claude Instant). There 406 multiple-choice, 13 true/false, and 4 "choose N" covering topics surgery, pediatrics, other disciplines. accuracy the output graded. Statistics used analyze differences among LLMs. Leaked excluded primary analysis. 4.0 scored (78.2%), Bing (67.2%), (64.4%), Instant (62.9%). Perplexity lowest (56.1%). Scores differed significantly between overall ( p < 0.001) pairwise comparisons. All higher multiple-choice vs true/false or “choose N” demonstrated limitations answering certain questions, indicating refinements needed before reliance education. expanding capabilities suggest a improve training if thoughtfully implemented. Further research should explore specialty specific optimal integration into curricula.

Language: Английский

---