Background/Objectives: Alzheimer's disease (AD) is a progressive neurodegenerative disorder necessitating accurate and timely diagnosis for effective clinical intervention. While deep learning methods have shown promise in AD classification, many rely on computationally intensive architectures and high-dimensional feature representations. This study introduces a lightweight hybrid framework combining deep feature extraction, dimensionality reduction, and adaptive classification for MRI-based Alzheimer's disease stage classification. Methods: Utilizing MRI images from a publicly available Alzheimer's disease dataset encompassing four clinical stages (Non-Demented, Very Mild Demented, Mild Demented, and Moderate Demented), deep features were extracted using a pre-trained SqueezeNet model as a fixed feature extractor, generating 1000-dimensional feature vectors. Due to the computational complexity and for the improvement of the model efficiency, the dimensionality reduction technique, Principal Component Analysis (PCA) was then applied. This resulted in an optimum representation of 100 principal components, retaining about 96% of the variance. Then, the performances of various machine learning classifiers such as k-Nearest Neighbors (kNN), Support Vector Machine (SVM), Decision Tree (DT), Neural Network (NN), Naïve Bayes (NB), Logistic Regression (LR) and Enhanced Fuzzy Min-Max Neural Network (EFMM) were tested. The accuracy, precision, recall, F1-score, area under the receiver operating characteristic curve (AUC), and confusion matrices were used to evaluate the performance. Stratified 5-fold cross validation was used to ensure the strength of our results. Results: The findings show that PCA has a significant improvement in classification accuracy for most of the models. In particular, the EFMM classifier outperformed the other classifiers, with an accuracy of 97.19% on the independent test set. After PCA, the AUC values for classes such as Mild Demented, Moderate Demented, Non-Demented and Very Mild Demented were obtained as 97.12%, 99.97%, 93.79% and 95.26% respectively. We further validated our proposed framework using stratified 5-fold cross validation which further corroborated the robustness of our proposed framework. The EFMM achieved a mean accuracy of 98.38% ± 0.36 and a mean macro-F1 score of 98.48% ± 0.43. Friedman statistical testing demonstrated that there were significant differences between the performance of the classifiers evaluated (p < 0.001), which further validated the performance of the EFMM. Conclusions: To sum up, the proposed SqueezeNet-PCA-EFMM is an effective and efficient method for Alzheimer's disease stage classification under MRI images. The combination of SqueezeNet, PCA, and EFMM-led not only to high classification performance, but also to good cross validation results. Furthermore, this property of incremental learning is the intrinsic one of the EFMM and renders this framework interesting for its incorporation in the next-generation intelligent clinical decision supports in particular, as medical care evolves.
Background: High-resolution ultrasound is increasingly used as a complementary tool in carpal tunnel syndrome (CTS), but the value of median nerve cross-sectional area (CSA) for severity stratification remains uncertain. This study evaluated whether median nerve CSA measured at the pisiform level can distinguish severe from moderate CTS classified according to the Bland grading scale. Methods: This prospective cross-sectional study included 72 participants: moderate CTS (n = 34), severe CTS (n = 24), and healthy controls (n = 14). Nerve conduction studies, abductor pollicis brevis needle electromyography, and ultrasound examinations were performed on the same day. CSA was measured three times by a blinded examiner, and the mean value was used for analysis. Correlation, age-adjusted linear regression, and receiver operating characteristic curve analyses were performed. Results: Median nerve CSA increased progressively from controls to moderate and severe CTS groups (8.0 ± 1.1, 16.5 ± 4.6, and 19.8 ± 3.4 mm2, respectively; p < 0.001), with significant pairwise differences. Intraobserver and interobserver reliability were excellent (ICC = 0.95 and 0.91, respectively). CSA correlated positively with distal motor and sensory latencies and inversely with compound muscle action potential and sensory nerve action potential amplitudes. CSA also showed a weak positive association with neurogenic motor unit action potentials. In age-adjusted regression analysis, distal motor latency remained independently associated with CSA (B = 0.873, p = 0.002). A CSA cut-off of 18.5 mm2 distinguished severe from moderate CTS with an area under the curve (AUC) of 0.728, sensitivity of 66.67%, and specificity of 79.41%. Conclusions: Median nerve CSA measured at the pisiform level may provide useful adjunctive structural information for CTS severity assessment and may show moderate discriminatory performance for differentiating severe from moderate CTS. CSA should be interpreted in conjunction with clinical and electrodiagnostic findings rather than as a stand-alone severity marker.
Background/Objectives: Accurate peripheral blood cell detection remains difficult in dense smear fields, where touching erythrocytes, small platelets, and staining variability reduce the reliability of box-based object detectors. This study aims to address these limitations by proposing BloodContourNet, a contour-informed and density-aware detection framework for red blood cell, white blood cell, and platelet localization. Method: BloodContourNet is built on YOLOv26-L and integrates three complementary modules. The Hematological Attention Module (HAM) refines high-level features using local convolution and shifted-window attention. The Density-Aware Reweighting Head (DARH) modulates the classification loss for locally sparse or difficult instances without image-level resampling. The Bézier Contour Refinement Head (BCRH) provides a weakly supervised geometric overlap cue for ambiguous dense-region post-processing. Results: Experiments on the TXL-PBC dataset, containing 1260 peripheral smear images and 18,143 annotated cells, benchmarked 25 YOLO variants and 10 Transformer-based or reference detectors under a unified split and evaluation protocol. Across five random seeds, BloodContourNet achieved 99.1% mean Average Precision (mAP) at 50% Intersection-over-Union (IoU) and 87.8% mAP@50:95, improving strict-IoU performance by 1.4 points over YOLOv26-L and 0.7 points over DINO-DETR. Both margins were statistically significant across the five seeds (paired two-sided t-tests, both p < 0.05), and the five-seed 95% confidence interval of the proposed model was [87.5, 88.0]. Conclusions: BloodContourNet primarily improved the strict-localization setting on TXL-PBC, where dense erythrocyte regions and small platelet instances remain challenging for box-based detectors. External patient- or slide-independent validation remains necessary before broader clinical claims can be made.
Background/Objectives: Artificial intelligence (AI) is increasingly being evaluated for ophthalmic diagnosis, screening, and triage, yet its role in paediatric eye care remains less established than in adult ophthalmology. This systematic review aimed to synthesise evidence on AI-enabled tools for paediatric ophthalmic diagnosis, screening, triage, surveillance, and referral, with an emphasis on diagnostic performance, safety, workflow integration, equity, and implementation readiness in primary, community, and primary care-relevant settings. Methods: A PRISMA-guided systematic review was conducted using MEDLINE, Embase, Web of Science, Scopus, and IEEE Xplore from inception to 30 March 2026. Eligible studies evaluated AI or machine-learning tools for children and adolescents aged 0-18 years in relation to paediatric eye conditions. Study selection and data extraction were undertaken independently by reviewers, with disagreements resolved by consensus or third-reviewer adjudication. Methodological and reporting quality was evaluated using an author-adapted six-domain rubric informed by APPRAISE-AI. Diagnostic-accuracy studies were assessed using an author-adapted QUADAS-2 framework incorporating QUADAS-AI-informed AI-specific considerations, the prediction-model study was assessed using PROBAST+AI, and the non-randomised treatment-effect study was assessed using ROBINS-I. The public dataset descriptor was evaluated separately using an author-developed dataset-quality, representativeness, and applicability framework. Because of clinical and methodological heterogeneity, findings were synthesised thematically. Results: Twelve empirical studies and one public dataset descriptor were included, covering retinopathy of prematurity, retinoblastoma, amblyopia risk, myopia, congenital cataract, and visual-acuity assessment. AI systems frequently demonstrated promising diagnostic or screening performance, including sensitivity-first detection of treatment-requiring retinopathy of prematurity, high discrimination for retinoblastoma activity, and strong myopia prediction using fundus images. Several studies supported feasibility in neonatal, school, and community workflows using smartphone-based imaging, task-shifted operators, tele-referral, and human-in-the-loop review. However, external and temporal validation, calibration, patient-level reporting, subgroup and fairness assessment, and economic evaluation were limited. Conclusions: AI-enabled tools show promise for supporting selected paediatric ophthalmic screening, triage, and surveillance pathways, particularly when combined with image-quality control, explicit escalation, and human oversight. However, confidence in the reported performance is limited by single-centre studies and enriched samples, small numbers of clinically important cases, heterogeneous analytical units, potentially optimistic aggregation procedures, limited external or temporal validation, incomplete calibration, and absent fairness analyses. Routine autonomous implementation remains premature.
Objectives: This study aims to compare the imaging performance of spectral-domain optical coherence tomography (SD-OCT) and swept-source OCT (SS-OCT) in the early postoperative assessment of patients undergoing vitreoretinal surgery with intraocular gas or air tamponade. Methods: Seventeen eyes of 17 patients who underwent pars plana vitrectomy with either sulfur hexafluoride (SF6, 20%) or air tamponade were prospectively enrolled. All patients underwent OCT imaging on postoperative day 1 using both the SD-OCT system and the SS-OCT platform, without pharmacological mydriasis. Images were independently evaluated by two experienced ophthalmologists based on the ability to delineate four prespecified anatomical layers: inner retinal layers, ellipsoid zone (EZ), retinal pigment epithelium (RPE) and choroid. Images were classified as adequate quality if two or more of these structures were identifiable. Results: SS-OCT provided adequate-quality images in all 17 eyes (100%), with complete visualization of the inner retinal layers, EZ, RPE, and choroid in 15 eyes (88.2%). In contrast, SD-OCT yielded adequate-quality images in only three of 17 eyes (17.6%), demonstrating marked signal attenuation, interface artifacts, and inability to resolve deeper retinal structures in most cases. No difference in imaging performance was observed between SF6 and air tamponade subgroups. Conclusions: SS-OCT demonstrates markedly superior imaging performance in gas-filled eyes on postoperative day 1 compared to SD-OCT, primarily attributable to its longer wavelength, reduced sensitivity roll-off, and superior penetration through optically challenging media. These findings suggest that SS-OCT may offer meaningful advantages for early postoperative monitoring following vitreoretinal surgery with tamponade; confirmation in larger prospective cohorts with clinical-outcome correlation is warranted before it can be recommended as the preferred modality.
