1 of 6

TEAM NAME : [YOUR TEAM NAME]

TEAM LEADER NAME : [TEAM LEADER NAME]

WHICH DOMAIN DOES YOUR IDEA ADDRESS : HealthTech · AI for preventable blindness

BAPUJI INSTITUTE OF ENGINEERING AND TECHNOLOGY

EXPLAINABLE AI FOR

EXPLAINABLE AI FOR

DIABETIC RETINOPATHY

DIABETIC RETINOPATHY

SCREENING IN RURAL INDIA

SCREENING IN RURAL INDIA

THE AI THAT SHOWS ITS WORK

Spots sight-threatening retinopathy early. Explains every call. A doctor signs off in 30 seconds.

RETINA SAYS:

GRADE 2

MODERATE DR

REFERABLE: YES

Heatmap shows where the AI looked.

ILLUSTRATIVE EXAMPLE · SYNTHETIC IMAGE

2 of 6

LEVEL 1: THE SILENT BLINDNESS

LVL 2/6 · WHAT PROBLEM ARE YOU SOLVING?

77M+

diabetic adults in India, 2nd highest in the world

~18%

of them develop diabetic retinopathy (DR)

1:100K

ophthalmologists per rural population

90%

of vision loss is preventable with early screening

BLACK-BOX AI

Gives a grade with no evidence. Doctors can't verify it, so they can't trust it.

UNVALIDATED

Tested on clean hospital images with little rigour. Unproven in real clinics.

BREAKS IN THE FIELD

Portable cameras give blurry, dark, off-centre photos that models never saw.

BOSS: TOO FEW DOCTORS. TOO MANY DIABETICS. NO TRUST IN BLACK BOXES.

EXPLAINABLE AI FOR DR SCREENING · MINEARENA 1.0 · ILLUSTRATIVE EXAMPLES ARE SYNTHETIC

LVL 2/6

3 of 6

POWER-UP: THE RETINA LOOP

LVL 3/6 · YOUR SOLUTION · HOW IS IT DIFFERENT?

STEP 01

CHECK

Quality gate rejects blurry, dark or off-centre shots with a recapture reason.

STEP 02

SEE

CLAHE rescues borderline photos. Vessel map shows the retina's structure.

STEP 03

GRADE

EfficientNet-B0 grades ICDR 0-4 and flags referable DR (grade 2+).

STEP 04

EXPLAIN

Heatmap + confidence + one-page report. Doctor verifies in under 30 s.

THE RULE: The AI screens. The doctor decides. Nothing is final without sign-off.

WHY IT'S DIFFERENT

BUILT FOR BAD FIELD PHOTOS

Quality gate runs before any diagnosis.

SHOWS EVIDENCE, NOT A VERDICT

Heatmap, confidence and report per case.

PLANNED FOR 100K+ / YEAR

Simulated queue sizes cameras and reviewers.

EXPLAINABLE AI FOR DR SCREENING · MINEARENA 1.0 · ILLUSTRATIVE EXAMPLES ARE SYNTHETIC

LVL 3/6

4 of 6

ACHIEVEMENTS UNLOCKED

LVL 4/6 · LIST OF FEATURES OFFERED

QUALITY GATE

Rejects blurry, dark or off-centre photos with a recapture reason. CLAHE rescues borderline ones.

WORKING

ICDR GRADING 0-4

EfficientNet-B0 fine-tuned on APTOS 2019. Flags referable DR (grade 2+) for referral.

WORKING

GRAD-CAM HEATMAP

Shows where the model looked, so a doctor can check the grade at a glance.

WORKING

30-SECOND REPORT

One page: grade, referral, heatmap and a concur / override box for the ophthalmologist.

WORKING

CONFIDENCE + VESSELS

Class probabilities, a low-confidence flag for manual review, and a Frangi vessel overlay.

BUILDING

CAPACITY SIMULATOR

Queue model sizes cameras, bandwidth and reviewers for 100,000+ patients a year.

NEXT

Black boxes TELL. This one SHOWS WHY.

EXPLAINABLE AI FOR DR SCREENING · MINEARENA 1.0 · ILLUSTRATIVE EXAMPLES ARE SYNTHETIC

LVL 4/6

5 of 6

THE WORKFLOW

LVL 5/6 · WORKFLOW / IDEA + TECH STACK

AT THE PHC

AI ENGINE · CLOUD / EDGE

TELE-REVIEW

CAPTURE

Portable fundus camera

QUALITY

OK?

ENHANCE

CLAHE + vessels

GRADE

EfficientNet-B0 ICDR 0-4

EXPLAIN

Grad-CAM + report

DOCTOR

Concur or override in <30 s

YES

NO: RECAPTURE + REASON

REFER

ROUTINE

INVENTORY · TECH STACK

PYTHON · PYTORCH

TIMM · EFFICIENTNET-B0

OPENCV · SCIKIT-IMAGE

PYTORCH-GRAD-CAM

GRADIO · HF SPACES

APTOS 2019 · IDRiD

SIMPY (SIMULATION)

KAGGLE · GITHUB

EXPLAINABLE AI FOR DR SCREENING · MINEARENA 1.0 · ILLUSTRATIVE EXAMPLES ARE SYNTHETIC

LVL 5/6

6 of 6

THE ARCHITECTURE

LVL 6/6 · HIGH-LEVEL ARCHITECTURE / USE-CASE DIAGRAM

HEALTH WORKER

Uploads photo, gets result or recapture tip

OPHTHALMOLOGIST

Reviews report, concurs or overrides

DR SCREENING ENGINE

1 QUALITY GATE

focus · light · field of view

2 ENHANCE

CLAHE + denoise

3 SEGMENT

vessel map (Frangi)

4 GRADE

EfficientNet-B0 ICDR 0-4

5 EXPLAIN

Grad-CAM + confidence

6 REPORT

1 page · sign-off box

DISTRICT OFFICER

Plans cameras and reviewers

CAPACITY SIMULATOR

SimPy queue: arrivals, bandwidth, reviewers

TRAINING DATA

APTOS 2019 · IDRiD

MODEL FILE

best_dr_model.pth

AUDIT LOG

grades + overrides

LEGEND

→ data flow

→ recapture / planning

→ models + logs

EXPLAINABLE AI FOR DR SCREENING · MINEARENA 1.0 · ILLUSTRATIVE EXAMPLES ARE SYNTHETIC

LVL 6/6