From Theory to Deployment

AI Applications

Every DeepMathAI application starts as a peer-reviewed publication. We turn our mathematically grounded models into transparent, explainable tools that anyone can use.

Medical Imaging AILive

Glaucoma Multi-View Screening

AI-powered opportunistic glaucoma screening from retinal fundus images.

Single & Batch PredictionMulti-View FusionExplainable AI (Grad-CAM)Performance Dashboard

Siying, P., et al. & Wangkeeree, R. — Applied Sciences (MDPI), Vol. 16, Issue 7, 3158 (2026)

Glaucoma Multi-View Screening — research pipeline
Live · AI

Pipeline

Glaucoma suspected · Grad-CAM ready

How It Works

A web-based decision-support system that flags retinal fundus images as "glaucoma suspected" by fusing predictions from two complementary views of the same image — the full fundus photograph and an automatically localized crop around the optic disc. The multi-view ensemble weights the disc-focused view more heavily, mirroring how ophthalmologists examine the optic nerve head, and every prediction can be inspected through Grad-CAM attribution maps.

Models

  • EfficientNetV2-M (full image)
  • EfficientNetV2-S (optic-disc crop)
  • Weighted multi-view ensemble

Key Features

Single & Batch Prediction

Screen one image or an entire folder of fundus photographs (JPG, PNG, BMP, TIF) in a single run.

Multi-View Fusion

Combines a full-image EfficientNetV2-M and a disc-crop EfficientNetV2-S with weighted probability fusion (Original × 0.5 + Crop × 1.5).

Explainable AI (Grad-CAM)

Heatmaps reveal the retinal regions driving each prediction, supporting clinical trust and error analysis.

Performance Dashboard

Built-in metrics page reporting the model's validated screening performance.

Published Research

Multi-view machine learning with an optic disc localization for glaucoma diagnosis

Siying, P., et al. & Wangkeeree, R. — Applied Sciences (MDPI), Vol. 16, Issue 7, 3158 (2026)

Important limitation: Decision-support tool for screening only — not a medical diagnosis and not a substitute for examination by an ophthalmologist.

Medical Imaging AILive

VCF MRI2BMD — Osteoporosis Screening

Opportunistic osteoporosis screening from routine lumbar-spine MRI.

Anatomically Guided Pipeline546 Radiomic FeaturesSingle & Batch PredictionExplainable Results

Mahatthanatrakul, A., Klinsuwan, T., Wangkeeree, R., & Laoruengthana, A. — Diagnostics (MDPI), Vol. 16, 2241 (2026)

VCF MRI2BMD — Osteoporosis Screening — research pipeline
Live · AI

Pipeline

L1–L4 localization → radiomics → BMD assessment

How It Works

Upload a sagittal lumbar-spine MRI slice and the system automatically detects the L1–L4 vertebrae, segments each vertebral body, extracts the study's 546-column radiomic feature vector, and runs it through validated classification and regression models to flag possible osteoporosis and estimate bone mineral density (BMD) and T-score. Because it works on scans acquired for other clinical reasons, it enables early, opportunistic detection at no additional cost to the patient.

Models

  • Optimization-refined SVM classifier (BMD status)
  • Support Vector Regression for BMD / T-score
  • Neurodynamic-SVM (RBF, tanh) — 85.2% published accuracy

Key Features

Anatomically Guided Pipeline

Multi-stage SSIM localization plus mutual-information detection isolates L1–L4, with differential-evolution segmentation of each vertebral body.

546 Radiomic Features

Extracts GLCM, LBP, wavelet, Gabor, Hu-moment, and first-order statistical features per vertebra for a rich, interpretable descriptor.

Single & Batch Prediction

Screen an individual case or queue a batch; each prediction runs as a background job in roughly 10–35 seconds.

Explainable Results

A dedicated explainability view surfaces the features and model reasoning behind each screening decision.

Published Research

An Anatomically Guided and Optimization-Refined Radiomics Framework for Opportunistic Osteoporosis Assessment from Lumbar Spine MRI

Mahatthanatrakul, A., Klinsuwan, T., Wangkeeree, R., & Laoruengthana, A. — Diagnostics (MDPI), Vol. 16, 2241 (2026)

Important limitation: Research decision-support demo — not a diagnostic device. Every output must be confirmed by a radiologist or qualified clinician, and identifiable patient images should never be uploaded.

Medical Imaging AILive

QMG-RID Explorer — Fundus Image Enhancement

Compare fundus-image enhancement methods live in your browser and watch QMG-RID optimize iteration by iteration.

Runs Entirely In-BrowserSide-by-Side Method ComparisonWatch the Optimization UnfoldDiagnosis & Quality Panel

Wangkeeree, R., Klinsuwan, T., & Luangsawang, K. — manuscript under review (2026)

QMG-RID Explorer — Fundus Image Enhancement — research pipeline
Live · AI

Pipeline

Input → QMG-RID → enhanced output

How It Works

Upload a retinal fundus photograph (or pick an example) and compare enhancement methods side by side: our QMG-RID decomposition, gray-world correction, CLAHE, non-local-means denoising, and the total-variation method of Wang et al. (2021) are all computed live in JavaScript on your device, alongside a precomputed reference from the GAN-based Cofe-Net. QMG-RID splits the image into a base and a detail layer and optimizes them toward a closed-form, color-preserving illumination target and a denoised detail target with per-pixel Adam descent — and the explorer lets you scrub through that optimization step by step.

Models

  • QMG-RID — illumination-flattening + detail-preservation decomposition (per-pixel Adam)
  • Baselines: gray-world, CLAHE, non-local means, Wang et al. (2021) TV decomposition
  • Cofe-Net (Shen et al., 2021) — precomputed reference only
  • Frozen DINOv2 ViT-B/14 + 3-class SVM DR classifier — 71.2% 5-fold CV accuracy

Key Features

Runs Entirely In-Browser

Every live method is plain JavaScript running on your own device — uploaded photographs are never sent to a server.

Side-by-Side Method Comparison

Switch between QMG-RID, gray-world, CLAHE, NLM denoising, Wang et al. (2021), and a precomputed Cofe-Net reference, with per-channel (RGB, HSV, CIE L*) before/after views.

Watch the Optimization Unfold

Play or scrub QMG-RID's iterations, inspect its base and detail layers, and toggle each loss term off with ablation presets to see what it contributes.

Diagnosis & Quality Panel

For the example images, real precomputed DR-class predictions and image-quality metrics from the paper's Python pipeline, original vs. enhanced.

Research Manuscript

Image Quality and Enhancement in Diabetic Retinopathy Screening: A Matched-Domain, Cross-Validated Evaluation with Self-Supervised Representations

Wangkeeree, R., Klinsuwan, T., & Luangsawang, K. — manuscript under review (2026)

Important limitation: Research demo accompanying a manuscript under review — not a diagnostic device. The in-browser methods are disclosed, simplified reimplementations at a 160×160 working resolution, so on-screen metrics are illustrative rather than the paper's reported results.

Medical Imaging AIBeta

NU-EyeFAST

Fundus AI Screening Tool — lesion map, glaucoma and diabetic retinopathy assessment for each eye from a fundus photograph or EIDON report.

Reports or Photographs13-Class Lesion MapGlaucoma & DR AssessmentStructural Measurements

DeepMathAI research group — manuscript in preparation (2026)

NU-EyeFAST — research pipeline
Beta · AI

Pipeline

Fundus → 13-class lesion map → glaucoma & DR

How It Works

A clinician-facing decision-support tool built on models trained by pooling public and Naresuan University Hospital fundus datasets. For every eye found in an uploaded EIDON report (PDF or screenshot) or fundus photograph, it segments 13 structures and lesions with a partial-label U-Net, measures cup-to-disc ratios and neuroretinal rim widths against Naresuan reference ranges, and scores glaucoma suspect, any diabetic retinopathy, severe DR and diabetic macular oedema — each placed against validated rule-out and refer thresholds. Access is restricted to the research team and approved collaborators.

Models

  • Partial-label U-Net — 13-class segmentation of disc, cup and lesions
  • Glaucoma suspect: DINOv2-B fine-tuned at 448 px — AUROC 0.959 internal, 0.791 on SMDG-19
  • Any DR: lesion features + DINOv2-L hybrid — AUROC 0.972 internal, 0.930 on IDRiD
  • Severe DR: gradient-boosted trees on lesion features — AUROC 0.956 internal, 0.929 on IDRiD
  • DME: EfficientNet-B0 — AUROC 0.987 internal, 0.945 on IDRiD

Key Features

Reports or Photographs

Upload EIDON report PDFs, screenshots of report pages, or plain fundus photographs from any camera; each fundus is located and analysed separately. Patient text on the report is never read.

13-Class Lesion Map

Optic disc, cup, DR lesions (microaneurysms, haemorrhages, exudates) and glaucoma-related signs (RNFL defects, notching, disc haemorrhage) shown as an interactive overlay.

Glaucoma & DR Assessment

Glaucoma suspect (fine-tuned DINOv2-B plus a structure-based model), any DR, severe DR and DME, each reported as Low / Review / Refer against Naresuan-validated operating points.

Structural Measurements

Vertical, horizontal and area cup-to-disc ratios, ISNT rim widths and the ISNT rule, with percentiles against Naresuan normal eyes.

Research Manuscript

Pooling Fundus Datasets for Lesion Segmentation, Glaucoma and Diabetic Retinopathy Screening

DeepMathAI research group — manuscript in preparation (2026)

Important limitation: Research prototype — not a medical device. Every result must be confirmed by a clinician. Thresholds were set on Naresuan validation images and should be re-set on locally graded images before use with another camera. Login required: access is limited to the research team and approved collaborators.

Medical Imaging AIBeta

SpineBMD Workstation

Low-BMD triage from routine lumbar-spine MRI — 3D segmentation, level-checked measurements and an explained T-score estimate.

Whole-Study DICOM Upload2D + 3D Segmentation ReviewLevel-Checked, Degeneration-AwareExplained Result & Report

Klinsuwan, T., et al. & Wangkeeree, R. — manuscript in preparation (2026)

SpineBMD Workstation — research pipeline
Beta · AI

Pipeline

DICOM → 3D segmentation → explained T-score

How It Works

A clinician-facing research workstation that flags low bone mineral density (osteopenia or osteoporosis) on routine sagittal lumbar-spine MRI. Doctors upload the DICOM study; the system picks the sagittal T1, T2 and STIR series, segments the vertebrae, discs, spinal canal and paraspinal muscles, checks the vertebral level labels across sequences, measures the L1–L4 vertebral bodies (signal ratios, texture radiomics, morphology and degeneration markers), and returns an estimated T-score with the exact contribution of every model input. Labels can be corrected in a linked 2D/3D viewer and the result exported as a PDF report. Access is restricted to the research team and approved collaborators.

Models

  • TotalSegmentator (vertebrae_mr + total_mr) — vertebra, disc, canal and muscle segmentation
  • Elastic-net model on 24 of 369 signal-ratio, radiomic, morphology and degeneration features
  • Site standardisation + Platt scaling — AUC 0.88 (NU) and 0.71 (BHUD); 0.86 / 0.67 when the hospital was unseen in training

Key Features

Whole-Study DICOM Upload

Drop files, folders or a zip with one or many patients; studies are grouped automatically and the sagittal T1, T2 and STIR series are chosen by the study's selection rule.

2D + 3D Segmentation Review

Sagittal slice viewer with overlay, linked to a 3D model of vertebrae, discs, canal and muscles. Relabel or exclude a vertebra and the result is recalculated.

Level-Checked, Degeneration-Aware

Automatic level-labelling QC across sequences, and a warning when degenerative change may make spine DXA read falsely high.

Explained Result & Report

Estimated T-score with its typical error and the exact (linear SHAP) contribution of each input; export a PDF report and a measurement CSV.

Research Manuscript

Detecting Low Bone Mineral Density on Routine Lumbar Spine MRI: A Three-Centre Study of Interpretable, Degeneration-Aware Models with Leave-One-Site-Out Validation

Klinsuwan, T., et al. & Wangkeeree, R. — manuscript in preparation (2026)

Important limitation: Research prototype — not a medical device and not a replacement for DXA. Every result must be confirmed by a clinician, and the threshold must be calibrated on about 50 local examinations with DXA before use at another hospital. Login required: access is limited to the research team and approved collaborators, and the GPU server is started on request.

More Applications in Development

Upcoming tools from our research pipeline include diabetic retinopathy screening, vertebral fracture risk stratification, and renewable-energy optimization systems. Follow our publications to see what's next.

Research You Can Actually Use

Interested in collaborating on deployable, explainable AI for healthcare or energy? DeepMathAI welcomes research partnerships and pilot deployments.