Marcos Antônio Dórea Machado

Background: Artificial intelligence (AI) may improve non-invasive liver fibrosis assessment, but its performance across different liver disease spectra remains uncertain. We evaluated laboratory-based AI models across two clinically distinct cohorts from a single institution and quantified their incremental performance over FIB-4. Methods: Two independent cohorts were analyzed: MASLD ( n = 135) and heterogeneous chronic liver disease (HEPATO; n = 126). Models using six routine laboratory biomarkers were evaluated through reciprocal cross-cohort validation and pooled three-fold cross-validation. Discrimination was compared with FIB-4 using paired DeLong tests, and classifications at approximately 80% and 90% sensitivity using McNemar tests. Six eligible literature studies informed a random-effects meta-analysis, Bayesian evidence synthesis, and clinical operating-point simulations. Results: Naive Bayes achieved AUCs of 0.76 in HEPATO and 0.75 in MASLD, compared with 0.68 and 0.72 for FIB-4, respectively; AUC differences were not statistically significant. In reciprocal cross-cohort operating-point analyses, AI showed numerically higher specificity than FIB-4 across all four validation scenarios, although achieved sensitivities differed between models, limiting direct specificity comparisons. Exploratory in-sample analyses suggested potential specificity gains requiring independent validation. The literature-only pooled Δ AUC was +0.074 (95% CI, +0.008 to +0.141; I ² = 82.3%). Bayesian updating yielded a posterior mean Δ AUC of +0.063 (95% CrI, approximately +0.0001 to +0.125), with an approximately 84.7% probability of Δ AUC >0.03. Simulated specificity gains were +10.9 and +10.2 percentage points at 80% and 90% sensitivity, respectively. Conclusions: Laboratory-based AI models demonstrated relatively consistent discrimination across two clinically distinct liver disease cohorts within a single-center setting. AI showed numerically higher specificity than FIB-4 across reciprocal cross-cohort operating-point analyses, although differences in achieved sensitivity preclude conclusions of superiority at matched operating points.

Background/Objectives: Artificial intelligence (AI) is an emerging technology for diagnosing liver fibrosis in Metabolic-Dysfunction-Associated Steatotic Liver Disease (MASLD), but a comprehensive synthesis of its performance is lacking. This systematic review (SR) aimed to evaluate the current evidence of AI models for diagnosing or staging liver fibrosis in patients with MASLD compared to conventional diagnostic tools. Methods: A comprehensive search was conducted in PubMed, Scopus, Web of Science, ScienceDirect, Embase, LILACS, IEEE Series, and Association for Computing Machinery (ACM). Primary studies applying AI to diagnose fibrosis in adults with MASLD were included. Risk of bias was assessed using the QUADAS-2 tool, and methodological reporting was evaluated according to the MINimum Information for Medical AI Reporting (MINIMAR) guideline. A narrative synthesis was performed, grouping studies by data type (clinical/laboratory vs. imaging) and summarizing diagnostic performance and clinical application. A frequency-based analysis was applied to identify the most recurrent predictive features, and an analysis of the AI architecture and application was reported. The review was registered in PROSPERO (CRD420251035919). Results: Twenty-one studies were included, encompassing 19,221 patients and 5237 images. Across studies, AI models consistently outperformed non-invasive scores such as Fibrosis-4 Index (FIB-4) and NAFLD Fibrosis Score (NFS). The most frequent predictive variables were identified. Despite an overall low risk of bias, methodological transparency and external validation were limited. Conclusions: AI is feasible for the non-invasive diagnosis of liver fibrosis in MASLD, demonstrating superior accuracy to standard clinical scores. Broader clinical application is limited by the lack of external validation and high heterogeneity among the studies. Prospective validation in diverse, multicenter cohorts is essential before AI can be integrated into routine clinical practice.

Extracting structured data from electronic health records (EHRs) remains a major challenge, particularly in non-English and resource-constrained healthcare systems. This study benchmarks multiple large language models (LLMs) for the automated extraction of structured clinical variables from Portuguese-language medical notes under limited computational resources. We evaluated five LLMs (GPT-4o mini, DeepSeek-V3, Mixtral-8x7B, LLaMA 8B, and Qwen-32B) against a manually curated dataset of cardiology and infectiology outpatient records. Models were deployed in quantized versions to optimize computational efficiency. Outputs were compared with human annotations using F1 score, balanced accuracy, and recall. Among the tested models, Qwen-32B achieved the highest performance in both the infectiology domain (balanced accuracy = 0.91 [0.07]) and cardiology domain (balanced accuracy = 0.89 [0.07]). Performance varied by clinical variable, with better results for frequently and consistently documented conditions (e.g., diabetes) and lower accuracy for complex or infrequent variables (e.g., tumors). Extraction time ranged from 0.9 to 24.2 minutes per patient, depending on clinical domain and model. These findings support the feasibility of applying LLM-based natural language processing tools in resource-limited, non-English healthcare settings. Future research should assess emerging high-parameter models and explore additional clinical domains.

Purpose: To develop and evaluate a deep learning (DL) framework for PET image denoising using scanner-specific noise power spectrum (NPS) simulations, enabling high-quality clinical reconstruction from low-dose or shortened acquisitions. Methods: A perceptually optimized 2.5D U-Net architecture was trained on paired clinical-clean and synthetically noised PET images, where the noise was generated by injecting NPS-derived correlated noise into clinical PET datasets from four scanners. Models were trained using datasets simulated with NPS (DL-NPS), list-mode reconstructions (DL-LM), or both (DL-all). The NPS dataset was augmented and hyperparameters were tuned until denoising performance reached optimal quantitative metrics. Image quality was assessed using SSIM, PSNR, RMSE, liver coefficient of variation (CV), and lesion contrast-to-noise ratio (CNR). Clinical detectability was analyzed using the Rose Criterion and Cochran’s Q test. SUV agreement was assessed via Bland-Altman analysis. Results: DL-NPS consistently improved image quality across scanners, performing competitively relative to DL-LM. DL-all further improved image quality for the EARL 2 setup, while DL-NPS remained the best performing model for the EARL 1 setup. At 25% and 50% counts, DL denoising improved CNR by 15–40% across systems, improving lesion detectability by up to 18%. Liver CV dropped below clinical thresholds post-denoising. Bland–Altman analysis showed minimal SUV bias (mean bias Conclusion: NPS-based simulation enables realistic, scanner-specific noise generation as a surrogate to scarce gated or list-mode data. The proposed DL approach consistently restores image quality and lesion detectability across scanners and imaging settings. Its implementation supports scalable clinical adoption of low-dose PET protocols.

Objective: Accessing structured data from electronic health records (EHRs) remains a major challenge, particularly in non-English and resource-constrained healthcare systems. This study benchmarks multiple large language models (LLMs) for automated extraction of structured clinical variables from Portuguese-language medical notes in a limited computational environment. Methods : We evaluated five LLMs (GPT-4o mini, DeepSeek-V3, Mixtral-8x7B, LLaMA 8B, and Qwen-32B) against a manually curated dataset comprising cardiology and infectiology outpatient records. Models were deployed in quantized versions to optimize computational efficiency. Model outputs were compared against human annotations using F1-score, balanced accuracy, and recall. Results : Among the tested models, GPT-4o mini consistently achieved the highest performance in the infectiology domain (balanced accuracy = 0.85 ± 0.06), while Qwen-32B performed competitively in cardiology (balanced accuracy = 0.86 ± 0.10). Model performance varied by clinical variable, with better results for frequently documented conditions (e.g., hypertension) and lower accuracy for complex or infrequent variables (e.g., tumors). Extraction time varied from 0.08 to 7.1 minutes per patient, depending on clinical domain and LLM model. Conclusion : These findings support the feasibility of deploying LLM-based natural language processing tools in resource-limited, non-English healthcare settings. Future research should assess emerging high-parameter models and explore additional clinical domains. Keywords : Large Language Models, Electronic Health Records, Clinical Data Extraction, Healthcare Digital Infrastructure.

Purpose: Current bone-scintigraphy protocols demand full-count, 10–15 min scans to preserve image quality, and existing deep-learning (DL) denoisers must be retrained for each camera manufacturer. We introduce a scanner-agnostic adaptive-diffusion U-Net that yields diagnostic-grade images from half-time or half-dose acquisitions without scanner-specific retraining. Methods: A multi-institutional retrospective set of 3,635 studies from four gamma-camera models was Poisson-thinned to 10–70% counts and partitioned for training/validation/testing. The network couples a U-Net backbone to a trainable isotropic diffusion layer. Quantitative evaluation (SSIM, PSNR, LPIPS) used 180 test scans. Prospective validation employed 60 new scans—including 20 from an unseen Philips camera—rated by three blinded nuclear physicians on a 5-point Likert scale. Head-to-head comparisons evaluated the diagnostic fidelity of DL approach. Results: At 50% counts, reconstructions raised SSIM from 0.87 ± 0.03 to 0.97 ± 0.02 and PSNR from 28.0 ± 4.0 dB to 39.0 ± 4.0 dB while lowering LPIPS from 0.07 ± 0.02 to 0.04 ± 0.02 (p < 0.01). Reader studies confirmed diagnostic equivalence between double-speed DL images and standard acquisitions (Likert: 4.4 ± 0.7 vs 4.5 ± 0.5). Conclusion: The adaptive-diffusion U-Net halves scan time and/or administered activity, preserves diagnostic integrity across scanners, and eliminates the need for scanner-specific retraining, enabling scalable, radiation-sparing clinical adoption. Key words : Bone scintigraphy; dose reduction; image denoising; adaptive U-Net architecture; scanner-agnostic AI

Purpose: Hodgkin Lymphoma staging and prognostic evaluation through radiomic and machine learning (ML) methods has been little explored. This study explores radiomic features and proposes a new ML radiomic model to predict treatment failure in Hodgkin lymphoma patients from PET images. Methods: 13 radiomic features were extracted from PET images of 51 subjects with Hodgkin’s Lymphoma with response to therapy and 6 subjects with treatment failure as classified by the Lugano criteria for lymphoma response. Univariate analysis was performed for feature selection by means of the areas under ROC curves (AUC), and Pearson correlation was assessed to reduce redundancy in the feature space. A ML algorithm was developed and trained-validated with a six-fold stratified cross validation. The ML performance was assessed through the results of 15 combinations of two-fold validation samples, deriving two ML methods (M1 and M2) for tumor classification at the tumor and patient levels. Results: The features zone percentage (ZP), high intensity large area emphasis (HILAE), entropy and standardized uptake value of maximum pixel (SUVmax) were independent predictors of treatment failure. The ML algorithm performed with AUC=0.86 (M1) and AUC=0.96 (M2) to classify “individual tumors” and at “patient level” performed with sensitivity = 80.0% /specificity = 88.3% (M1) and sensitivity = 100% /specificity = 100% (M2). The ML outperformed the Ann Arbor staging that achieved, respectively, 83.3% and 31.7% for sensitivity and specificity. Conclusions: Evaluation of baseline 18F-FDG PET scans of individuals with Hodgkin’s lymphoma using a ML radiomic model enabled accurate classification of patients at higher risk of treatment failure.

Purpose To evaluate the classification performance of structured report features, radiomics, and machine learning (ML) models to differentiate between Coronavirus Disease 2019 (COVID-19) and other types of pneumonia using chest computed tomography (CT) scans. Methods Sixty-four COVID-19 subjects and 64 subjects with non-COVID-19 pneumonia were selected. The data was split into two independent cohorts: one for the structured report, radiomic feature selection and model building (n = 73), and another for model validation (n = 55). Physicians performed readings with and without machine learning support. The model's sensitivity and specificity were calculated, and inter-rater reliability was assessed using Cohen’s Kappa agreement coefficient. Results Physicians performed with mean sensitivity and specificity of 83.4 and 64.3%, respectively. When assisted with machine learning, the mean sensitivity and specificity increased to 87.1 and 91.1%, respectively. In addition, machine learning improved the inter-rater reliability from moderate to substantial. Conclusion Integrating structured reports and radiomics promises assisted classification of COVID-19 in CT chest scans.

O advento da radiômica tem gerado oportunidades em várias aplicações diagnósticas, incluindo a tomografia computadorizada (TC) de tórax para diagnóstico e estadiamento da COVID-19. Este trabalho demonstra a aplicação de técnicas de seleção e harmonização de atributos radiômicos em um estudo multicêntrico de COVID-19. 37 indivíduos com COVID-19 e 36 indivíduos com outras pneumonias não associadas a COVID-19 foram selecionados de três centros diagnósticos. Noventa e quatro atributos radiômicos da TC de tórax foram quantificados e selecionados usando dois métodos: 1) supervisionado, utilizando análise univariada pela curva ROC, e 2) não supervisionado, utilizando concordância de observadores intraclasse (ICC > 0,9) a partir de réplicas das TC com variações na qualidade de imagem e na segmentação. Adicionalmente, foram analisadas características populacionais que pudessem enviesar a harmonização dos atributos. O método de seleção não supervisionado demonstrou ser eficaz na triagem de atributos harmonizados, não sendo necessário uma harmonização com pós processamento. Os atributos selecionados pelo método supervisionado possuem maior dispersão das distribuições das quantificações entre os equipamentos, fazendo-se necessária a aplicação da harmonização ComBat após avaliação da sensibilidade do viés ecológico.