Automated Diabetic Screening via Anterior Segment Ocular Imaging: A Deep Learning and Explainable AI Approach

arXiv — cs.CV•Tuesday, March 17, 2026 at 4:00:00 AM

PositiveArtificial Intelligence

What Happened
A new deep learning system has been developed for automated diabetic retinopathy screening using anterior segment ocular imaging, which is more accessible than traditional fundus photography. This system utilizes visible biomarkers in the iris, sclera, and conjunctiva to classify diabetic status, validated on 2,640 clinically annotated images.
Why It Matters
This advancement is significant as it addresses the challenges of limited access to specialized equipment in primary care settings, potentially improving early detection and management of diabetic retinopathy.
The Bigger Picture
The integration of deep learning techniques, particularly Vision Transformers and other architectures, highlights a growing trend in medical imaging where AI is increasingly used to enhance diagnostic accuracy and efficiency across various conditions, including other diseases like breast cancer and placenta accreta spectrum.

— via World Pulse Now AI Editorial System

Read Original

Was this article worth reading? Share it

One More Thing in AI

Master AI with curated tools and tutorials for practical, real-world applications.

LucidQuery AI

Combines diffusion reasoning with autoregressive LLM for advanced AI analysis.

AI & DataView app details

Attentive AI

Extract digital maps from satellite, aerial, and drone imagery using deep learning.

AI & DataView app details

The Visualizer

Transform complex topics into clear, visual explanations for effortless learning.

AI & DataView app details

Eyeye

Train your eyesight with real-time eye tracking and personalized exercises.

Tech & Developer ToolsView app details

Novaheadshot

Transform selfies into professional headshots with AI—no photographer required.

AI & DataView app details

Continue Readings

arXiv — cs.LGa day ago

Diabetic Retinopathy Classification using Downscaling Algorithms and Deep Learning

PositiveArtificial Intelligence

A recent study has proposed a novel approach for classifying Diabetic Retinopathy (DR) using downscaling algorithms and deep learning techniques, specifically employing a Multi Channel Inception V3 architecture. The research amalgamates two datasets, Kaggle and the Indian Diabetic Retinopathy Image Dataset, to enhance the accuracy of DR classification across five severity stages.

Read full article

via arXiv — cs.LG

arXiv — cs.CVa day ago

What-Where Transformer: A Slot-Centric Visual Backbone for Concurrent Representation and Localization

PositiveArtificial Intelligence

The What-Where Transformer (WWT) has been introduced as a novel visual backbone designed to enhance concurrent representation and localization in image understanding tasks. This approach emphasizes a separation of 'what' and 'where' information, addressing the complexities of object discovery, detection, and segmentation, which are often more challenging than simple image classification.

Read full article

via arXiv — cs.CV

arXiv — cs.CVa day ago

Confidence-Guided Diffusion Augmentation for Enhanced Bangla Compound Character Recognition

PositiveArtificial Intelligence

A new framework for recognizing handwritten Bangla compound characters has been proposed, addressing challenges such as complex character structures and limited high-quality annotated data. This confidence-guided diffusion augmentation framework combines class-conditional diffusion modeling with classifier guidance to synthesize high-quality samples, enhancing recognition capabilities.

Read full article

via arXiv — cs.CV

arXiv — stat.ML2 days ago

Supervised Dimensionality Reduction Revisited: Why LDA on Frozen CNN Features Deserves a Second Look

NeutralArtificial Intelligence

A recent study revisits the use of Linear Discriminant Analysis (LDA) for supervised dimensionality reduction on frozen pretrained image representations, demonstrating its effectiveness in improving classification accuracy across various datasets, including CIFAR-100 and Tiny ImageNet. The evaluation involved ten dimensionality-reduction strategies applied to features from six different backbones, revealing that LDA outperformed full feature usage in 11 out of 12 configurations, achieving accuracy gains of up to 4.5 percentage points while significantly reducing dimensionality.

Read full article

via arXiv — stat.ML

arXiv — cs.CV2 days ago

Scaling Vision Models Does Not Consistently Improve Localisation-Based Explanation Quality

NeutralArtificial Intelligence

A recent study has shown that scaling vision models, specifically within the ResNet, DenseNet, and Vision Transformer families, does not consistently enhance the quality of localization-based explanations. The research evaluated 11 models across three image datasets, using various explainable AI methods and metrics to assess mask alignment. Findings indicate that increased architectural depth and parameter count do not correlate with improved explanation quality in most cases.

Read full article

via arXiv — cs.CV

arXiv — cs.CV2 days ago

MicroViTv2: Beyond the FLOPS for Edge Energy-Friendly Vision Transformers

PositiveArtificial Intelligence

The introduction of MicroViTv2 marks a significant advancement in the field of Vision Transformers, offering a lightweight model optimized for edge deployment while maintaining high accuracy and energy efficiency. This model builds on the original MicroViT and incorporates innovative techniques such as Reparameterized Patch Embedding and Single Depth-Wise Transposed Attention, achieving improved performance on benchmarks like ImageNet-1K and COCO.

Read full article

via arXiv — cs.CV

arXiv — cs.CV2 days ago

Spatial-Frequency Gated Swin Transformer for Remote Sensing Single-Image Super-Resolution

PositiveArtificial Intelligence

Researchers have introduced the Spatial-Frequency Gated Swin Transformer (SFG-SwinSR) for single-image super-resolution in remote sensing, enhancing the reconstruction of high-resolution images from low-resolution inputs while preserving fine spatial details. This model modifies existing Swin Transformer architectures by incorporating a specialized feed-forward network that differentiates between low-frequency structural content and high-frequency details.

Read full article

via arXiv — cs.CV

arXiv — cs.CV2 days ago

Developing a foundation model for high-resolution remote sensing data of the Netherlands

PositiveArtificial Intelligence

A foundation model has been developed utilizing 1.2m high-resolution satellite images of the Netherlands, combining Convolutional Neural Networks and Vision Transformers to capture both fine and large-scale landscape features. The model leverages temporal data to enhance learning from contextual information over time, improving representation and generalization with fewer labeled samples.

Read full article

via arXiv — cs.CV

Ready to build your own newsroom?

Subscribe to unlock a personalised feed, podcasts, newsletters, and notifications tailored to the topics you actually care about