Diabetic retinopathy (DR) remains one of the leading causes of preventable blindness globally, particularly among individuals with long-standing diabetes. Early detection through regular eye examinations is essential to prevent irreversible vision loss associated with advanced stages, such as proliferative diabetic retinopathy and diabetic macular oedema. Although screening programs have been successfully deployed in various healthcare systems, rising diabetes prevalence places a growing strain on medical infrastructure. As a result, there is a critical need for scalable, automated diagnostic tools. Recent advances in artificial intelligence, particularly deep learning using convolutional neural networks (CNNs), offer promising solutions for automated analysis of retinal images. These models have demonstrated high diagnostic performance in identifying DR stages and detecting macular oedema in imaging modalities like optical coherence tomography (OCT). Several AI algorithms have now received regulatory approval and are gradually being adopted in clinical workflows. Furthermore, innovations in portable imaging devices open new avenues for patient-led monitoring and remote diagnostics. However, despite their potential, current mobile imaging systems often fall short in achieving the resolution and consistency required for reliable DR detection when compared to standard fundus photography. Integration into telemedicine platforms could bridge this gap by enabling remote screening and centralized analysis, yet real-world implementation remains limited. Challenges such as legal regulations, software interoperability, and misalignment with existing national screening protocols continue to hinder widespread adoption. This paper explores the current state of AI-assisted diabetic retinopathy screening, evaluates the readiness of emerging technologies, and discusses key barriers that must be addressed to enable global deployment and improve patient outcomes.
4 October 2025 in British Journal of Nursing Studies
Breast cancer is the most commonly diagnosed malignancy and a leading cause of cancer-related mortality among women in the United States (American Cancer Society, 2024). Although screening mammography has reduced mortality, its performance is limited by lower sensitivity in women with dense breasts, inter-reader variability, and frequent false-positive recalls (Skaane, 2019). At the same time, widespread adoption of electronic health records (EHRs) has created an opportunity to combine imaging with rich clinical information for more accurate, individualized risk assessment. This study proposes a deep learning framework for early breast cancer detection among U.S. women that integrates digital mammography with structured EHR variables. Mammography images are obtained from a curated public dataset such as the Curated Breast Imaging Subset of the Digital Database for Screening Mammography (CBIS-DDSM), which provides biopsy-verified benign and malignant cases (Lee et al., 2017). EHR-style data are organized following the structure of large U.S. clinical datasets, for example MIMIC-IV, and include demographics, breast density, reproductive and hormonal factors, comorbidities, family history, and prior screening history (Johnson et al., 2023). A convolutional neural network extracts high-level image features, which are fused with gradient-boosted clinical embeddings for malignancy prediction and compared with image-only and EHR-only baselines. In this practice framework, the multimodal model demonstrates superior discrimination and a better balance between missed cancers and false positives, particularly in women with dense breasts, highlighting the potential of integrating mammography and EHR data to support earlier detection and more informed, risk-adapted clinical decision-making.
Pneumonia remains a major threat to under-five mortality, especially where radiology access is limited and delayed reads are common. We introduce Pneunet, a compact convolutional network for binary classification of pediatric chest radiographs (pneumonia vs. normal). The design blends multibranch dilated convolutions for varied receptive fields, contextual pooling to preserve global cues, and Efficient Channel Attention to recalibrate salient channels. We trained and evaluated on 5,863 anterior-posterior pediatric CXRs from the public corpus released by Guangzhou Women and Children's Medical Center after quality screening, resizing, normalization, and data augmentation. On the held-out test set, Pneunet achieved 98.3% accuracy, 0.985 precision, 0.993 recall, $0.989 \mathrm{F}_{1}, 0.957$ specificity, and ROC-AUC of 1.00, outperforming VGG16 and ResNet50 under identical splits. The gains stem from multi-scale feature learning and channel attention that sharpen subtle pneumonic patterns. These findings suggest Pneunet is well-suited for integration into computer-aided diagnosis workflows in pediatric care, particularly in resource-constrained settings.
21 October 2026 in Discover Artificial Intelligence
Medical image classifiers are often evaluated by discrimination metrics, although their probability estimates may fail on data from a different imaging source. This study evaluates probability calibration in one cross-dataset retinal image experiment and tests limited-label target recalibration without model retraining. Five EfficientNet-B0 classifiers were trained on APTOS 2019 and evaluated on a disjoint Messidor-2 test set. Internal APTOS AUC was 0.997 ± 0.002, whereas external Messidor-2 AUC was 0.665 ± 0.066. An audit found exact or confirmed near-identical images spanning the image-level APTOS splits; excluding the 27 affected source-test images left AUC essentially unchanged at 0.997 ± 0.002. Target calibration was repeated for 200 random samples at each size (25, 50, 100, and 200 labels) and compared across temperature, Platt, beta, and isotonic calibration. Temperature-scaling equal-mass ECE15 decreased from 0.108 at 25 labels to 0.085 at 200 labels; beta calibration reached 0.061 at 200 labels. At 200 labels, beta calibration achieved a mean Brier score of 0.221 and Brier skill score of 0.104 relative to a calibration-sample-prevalence predictor. The fixed 100-label temperature fit reduced high-confidence error rate (HCER) at 0.9 from 0.220 to 0.001, but coverage simultaneously decreased from 0.667 to 0.005. Risk–coverage curves were unchanged because temperature scaling preserves confidence ranking. Labeled target recalibration improved probability error but did not restore discrimination or remove classification errors. Confidence-threshold results therefore require coverage and conditional-error reporting.
Brain tumors represent a major medical concern, often leading to severe neurological complications or fatal outcomes if not diagnosed promptly. Early and accurate detection using MRI scans is vital for guiding effective treatment decisions. However, conventional diagnostic approaches are frequently labor-intensive and susceptible to operator errors, highlighting the necessity for automated and intelligent systems. This research introduces a novel deep learning architecture named MSDSCCNet (Multi-Scale Depthwise Separable Coupled Convolutional Network), specifically designed for multi-class brain tumor identification. The model utilizes a set of parallel convolutional pathways with distinct kernel sizes (11×11, 9×9, 7×7, and 5×5) to capture both local and contextual features. These outputs are fed into a sequential refinement block with progressively smaller convolutional layers and regularization through dropout. To evaluate its performance, the model was trained and validated on a curated dataset of 7,023 brain MRI scans sourced from Figshare, SARTAJ, and Br35H. The dataset spans four classes: glioma, meningioma, pituitary tumor, and healthy controls. Preprocessing and data augmentation techniques were employed to improve model generalization. MS-DSCCNet achieved a remarkable accuracy of 99.69%, with F1-scores surpassing 0.99 across all categories. Furthermore, the model delivered exceptional outcomes in ROC-AUC and Dice Similarity Coefficient (DSC) evaluations. When benchmarked against leading architectures such as ResNet50, VGG16, and MobileNetV2, the proposed model demonstrated superior efficiency, generalization, and precision. These findings affirm the promise of MS-DSCCNet as a practical and dependable tool for early-stage brain tumor detection.