ACADEMIC RESEARCH SYSTEM / 01
RetinaScan AI is an academic screening system for retinal fundus photographs. It checks image quality, estimates patterns associated with diabetic retinopathy, and shows where the model looked.
FIELD NOTE / 02
A fundus photograph is taken in seconds — in a clinic corridor, lights low, patient tired. Most frames come back ordinary. A few carry quiet changes, a microaneurysm here, a blot hemorrhage there, exudates gathering near the macula, that are easy to miss when the reader is rushed, fatigued, or simply human.
RetinaScan AI was built for that moment. It does not stand in for a clinician's judgement. It organizes attention: it confirms the photograph is readable, screens for patterns associated with diabetic retinopathy, and leaves a visible trace of where the model looked — so a qualified reader can decide what the image really says.
METHOD / 03
Every screening passes through four stations. Each one writes to the same JSON record, and each one can be inspected on its own.
STATION 01 · PRE-ANALYSIS
Before any model sees the image, OpenCV checks it. Frames that fail field coverage, illumination, or focus are rejected and logged — never silently interpreted.
cv2 · field-coverage · luminance-window · blur-variance → reject log
STATION 02 · INFERENCE
Accepted images are preprocessed and read at 224×224 by a calibrated EfficientNet-B0, producing an academic binary screen: patterns associated with diabetic retinopathy, or not.
torch · efficientnet_b0 · 224×224 · binary head
STATION 03 · INTERPRETATION
The raw score is mapped through calibration so the number behaves like a probability. Value, threshold, and model version are written into the record together.
calibrated p̂ · threshold logged · model-version pinned
STATION 04 · EXPLANATION
Gradients from the final convolutional block are weighted back onto the image to show where the model attended. An attention map — never proof of a lesion.
grad-cam · last conv block · upsampled to field
PLATE / 04
One photograph, two views: the image the camera made, and the map the model left behind. Drag the divider — or use arrow keys — when a report exists.
NO ACTIVE REPORT
Run a screening to see the original image and model attention map here.
Start a screeningEVIDENCE / 05
This page presents the project's actual architecture as editorial annotations. No metric, confidence value, or status is displayed unless it comes from a verified artifact on this machine.
MODEL NOT READY — ARTIFACTS ABSENT
A single classification head: patterns associated with diabetic retinopathy, or not. Academic screening, not staging, not diagnosis.
Deterministic checks run before inference. Rejected frames are logged with their failure reason, keeping model errors interpretable.
Reported confidence is calibrated; raw logits are never presented as certainty, and the calibration version is pinned to each record.
Each screening writes one JSON record: image reference, quality result, calibrated probability, model version, Grad-CAM path.
A responsive React front end speaks to a FastAPI service — the same code path on a lab monitor or a phone in a corridor.
No database, no Docker, no PDF export, no fabricated prediction path. The constraints are part of the honesty of the demo.
No verified evaluation output is installed. Exact values will render here only from the file above — dataset, split, image size and training configuration included. Nothing is invented.
JOURNAL / 06
Short working notes on why the system is shaped the way it is.
NOTE 01 · QC FIRST
Most unreliable screening results begin upstream, in the photograph itself. Rejecting unusable frames with OpenCV — before any model sees them — keeps the record honest and the model's failures interpretable.
NOTE 02 · CALIBRATION
A raw score can rank images but cannot be read as risk. Calibration maps the classifier's output to a probability that behaves like one — and the mapping's version is written into every report.
NOTE 03 · GRAD-CAM
An attention map explains the model, not the retina. We publish it beside the same disclaimer every time, because the most dangerous overlay is the one that looks like a diagnosis.
KNOWLEDGE BASE / 07
Essential clinical context, technical methodologies, and screening classifications explained.
Diabetic Retinopathy is a microvascular complication of diabetes mellitus affecting the retina. Chronic high blood glucose damages retinal capillaries, leading to microaneurysms, intraretinal blot hemorrhages, hard lipid exudates, macular edema, and potential vision loss if left untreated.
RetinaScan AI applies a 4-stage pipeline: (1) An adaptive OpenCV image clarity gate with CLAHE microvascular enhancement; (2) A calibrated EfficientNet-B0 deep convolutional neural network for 4-class severity inference; (3) Grad-CAM gradient-weighted class activation heatmaps to highlight pathological attention; and (4) An institutional diagnostic report with print and email delivery.
The model grades fundus scans into four clinical stages: Stage 0 (No DR / Healthy Retina) with clear microvasculature; Stage 1 (Mild DR) characterized by early microaneurysms; Stage 2 (Moderate DR) marked by hemorrhages and hard exudates; and Stage 3 & 4 (Severe NPDR & Proliferative DR) indicating high-risk neovascularization.
The system accepts digital posterior-pole retinal fundus photographs (50° field) with a clear view of the macula, optic disc, and major vascular arcades in standard JPEG, PNG, or WebP formats (up to 10 MB).
BEGIN / 08
Upload a single fundus photograph. The system returns one record: quality result, calibrated probability, attention map — and the disclaimer that belongs with them.
USE FICTIONAL OR CONSENTED DATA · THIS ACADEMIC RESULT IS NOT A DIAGNOSIS.
