Automated mobile microscopy in Internet of Things (IoT) networks is essential for scaling malaria screening in resource-constrained environments. Deploying standard convolutional architectures here introduces severe adversarial vulnerabilities. Post-Training Quantization (PTQ) mitigates hardware constraints by converting floating-point models to 8-bit integers (INT8); however, its impact on clinical safety and security remains unexplored. This study presents an adversarial audit of quantized Vision Transformers for medical edge deployment. We evaluated a Swin-Tiny transformer against ViT-Tiny and MobileNetV3 baselines using a 27,558-image malaria dataset and an out-of-distribution (OOD) White Blood Cell dataset. Our findings redefine the “Quantization Shield” hypothesis. PTQ compresses the Swin model by 3.9× (to 27.89 MB) with a negligible 0.11% accuracy drop, maintaining statistical reliability on OOD tests. However, the hypothesized architectural resilience shatters under white-box Projected Gradient Descent (PGD) attacks. Despite robustness against single-step attacks, both MobileNetV3 and the INT8 Swin-Tiny collapse to 0.00% accuracy under iterative PGD. Conversely, the quantized Swin-Tiny resists black-box transfer attacks from a surrogate, maintaining 81.00% accuracy. We conclude that while quantized Vision Transformers meet mobile sensor constraints, integer quantization provides zero innate defense against targeted iterative perturbations, exposing a critical vulnerability in diagnostic IoT networks.

Background: Tricuspid valve dysfunction is historically underdiagnosed, and right-sided cardiac abnormalities are clinically important in high-risk diabetic populations. Diabetes may promote right-sided dysfunction through cardiometabolic remodeling, diastolic dysfunction, and elevated pulmonary pressures. This study introduces an explainable, lightweight artificial intelligence framework to infer extreme right heart phenotypes from left-sided echocardiographic and systemic clinical markers. Methods: We retrospectively analyzed an existing clinical dataset of approximately 370 Saudi Arabian individuals with diabetes mellitus. Seven baseline machine learning classifiers were evaluated using leak-aware preprocessing. To reduce optimism bias, models were validated with a True Nested Cross-Validation protocol. Probabilistic calibration, parameter reduction, and computational efficiency were prioritized for clinical triage, and SHapley Additive exPlanations (SHAP) supported transparent decision-making. Results: Random Forest was selected for its balance of discrimination and calibration. Under True Nested Cross-Validation, it achieved an AUC-ROC of 0.8602, AUC-PR of 0.8343, accuracy of 0.8075, sensitivity of 0.7733, specificity of 0.8239, Brier Score of 0.1281, Calibration Slope of 1.1720, and Observed/Expected Ratio of 0.8522. Feature ablation indicated a holistic cardiometabolic severity signal, with left atrial dimensions, diastolic dysfunction grade, and diuretic use as primary predictors. The model required 205.00 kilobytes of storage and 14.1956 milliseconds for inference. Conclusions: Left-sided cardiac markers combined with systemic indicators can flag concurrent severe right-sided abnormalities. This lightweight framework is a promising triage-oriented screening prototype for prioritizing echocardiographic assessment in high-risk diabetic patients during routine clinical visits without requiring specialized hardware; however, until prospective external validation is completed, this study must strictly be viewed as hypothesis-generating.

Background: Intelligent decision support systems for colonoscopy can fail when encountering out-of-distribution surgical instruments such as snares or biopsy forceps, producing false-positive polyp masks that may contribute to automation bias and reduce workflow reliability. This study aimed to develop and evaluate a parameter-efficient framework for open-set-aware polyp segmentation that can reject such anomalous inputs while preserving in-distribution segmentation performance. Methods: We propose PolypSAM-Open, which integrates a prototype-based Open-Set Learning (OSL) module with Low-Rank Adaptation (LoRA) in the MedSAM image encoder. The model was trained on Kvasir-SEG using an 85/15 split of authentic polyp images and synthetic high-frequency Gaussian noise to learn a rejection margin. Zero-shot out-of-distribution detection was evaluated on 590 unseen authentic surgical instruments from Kvasir-Instrument. Segmentation and detection performance were compared against a standard MedSAM-LoRA baseline. Results: Standard parameter-efficient fine-tuning yielded an OOD AUROC of 0.4263 on authentic surgical instruments. PolypSAM-Open improved zero-shot OOD AUROC to 0.9535 ($p<0.001$). Despite allocating 15% of training capacity to the synthetic-noise rejection margin, PolypSAM-Open maintained segmentation performance comparable to the standard fine-tuned baseline (Dice 0.9728 versus 0.9723). On ETIS-LaribPolypDB and CVC-ClinicDB, Dice scores were 0.9301 and 0.9386, respectively. The approach remained parameter-efficient, updating 4.48% of total parameters while adding negligible inference latency relative to the underlying MedSAM forward. Conclusions: Prototype-based open-set adaptation can substantially improve rejection of unseen surgical artifacts in AI-assisted colonoscopy while preserving high segmentation accuracy. These findings position PolypSAM-Open as a promising strategy for potentially safer decision-support segmentation in endoscopic workflows; prospective clinical validation remains necessary.

Objectives: The classification of Acute Lymphoblastic Leukemia (ALL) from peripheral blood smear images using Convolutional Neural Networks (CNNs) has achieved expert-level accuracy. However, the computational and memory requirements of CNNs pose a barrier to their deployment in resource-constrained clinical settings and low-income countries. To bridge this gap, we propose NeurALLNet, a memory-efficient convolutional spiking neural network (SNN) augmented with Squeeze-and-Excitation channel attention for the multi-class classification of ALL subtypes. Methods: NeurALLNet leverages sparse, event-driven temporal computation with an ultra-compact architecture of approximately 0.3M trainable parameters. The model was trained and evaluated on a primary dataset of ALL peripheral blood smear images, and its clinical generalizability was rigorously validated on an unseen external cohort of 3,242 images without retraining. We conducted hardware profiling on CPU and GPU platforms, alongside ablation studies and Grad-CAM visual explanations, to evaluate deployment viability and interpretability. Results: NeurALLNet achieved a test accuracy of 98.16% on the primary dataset, with a bootstrapped 95% Confidence Interval (CI) of [0.9663, 0.9939]. On the external validation cohort, it yielded an accuracy of 96.02%, with a robust 95% CI of [0.9534, 0.9667]. The architecture requires a memory footprint of 1.35 MB, achieving single-image inference latencies of 454.67 ms on a standard CPU and 11.24 ms on a GPU. Ablation studies confirmed that the attention mechanism is critical to the network’s discriminative power, and Grad-CAM visualizations verified that predictions are grounded in clinically relevant morphological features. Conclusion: Compared to recent state-of-the-art ensemble and hybrid CNNs that require millions of parameters, NeurALLNet delivers competitive diagnostic accuracy while reducing the computational footprint by orders of magnitude. By providing this precision within a 1.35 MB envelope, NeurALLNet offers a scalable, energy-efficient digital health intervention suitable for portable Lab-on-a-Chip devices and point-of-care diagnostics worldwide.