Ear Disease Detection

AI-Powered Ear Disease Detection with Explainable AI

A full-stack web application for analyzing otoscopic ear images using a PyTorch EfficientNet-B0 model. The system provides disease prediction, calibrated confidence, Grad-CAM visual explanations, AI-generated guidance, protected diagnosis history, and downloadable PDF reports.

Medical disclaimer: This project is an AI-assisted research/academic system. It is not a clinical diagnostic device and should not replace evaluation by a qualified healthcare professional.

---

Overview

The system supports five classes:

1. Acute Otitis Media
2. Cerumen Impaction
3. Chronic Otitis Media
4. Myringosclerosis
5. Normal

A typical prediction flow is:

Ear Image
   ↓
Flask API + Image Validation
   ↓
Image Quality Checks
   ↓
EfficientNet-B0
   ↓
Temperature-Scaled Probabilities
   ↓
Prediction + Confidence
   ├── Grad-CAM Heatmap
   ├── AI-generated Explanation
   └── Diagnosis History
            ↓
       PDF Report

---

Key Features

• EfficientNet-B0 image classification using PyTorch/Torchvision
• Transfer learning with staged fine-tuning
• Temperature scaling for confidence calibration
• Grad-CAM visual explanations
• Image validation for JPEG/PNG files
• Image resolution, blur, brightness, and contrast checks
• Maximum upload size of 50 MB
• Protected media access with ownership checks
• JWT authentication with access and refresh tokens
• Rate limiting on sensitive API endpoints
• Diagnosis history with individual/all-record deletion
• Downloadable professional PDF reports
• Model metadata endpoint
• Health monitoring endpoint
• React + Tailwind responsive frontend
• MongoDB persistence
• Automated backend tests with pytest
• GitHub Actions CI for backend tests and frontend builds

---

Model

EfficientNet-B0

Property	Value
Architecture	EfficientNet-B0
Framework	PyTorch / Torchvision
Input size	224 × 224
Classes	5
Parameters	4,013,953
Checkpoint	models/efficientnet_b0_ear_disease.pth
Model version	1.0.0

Inference preprocessing

Images are resized to 224 × 224, converted to tensors, and normalized using ImageNet statistics:

Mean = [0.485, 0.456, 0.406]
Std  = [0.229, 0.224, 0.225]

Training used image augmentation including horizontal flipping, rotation, and color jitter.

---

Model Comparison

Four pretrained CNN architectures were trained and evaluated using the same leakage-safe dataset and evaluation protocol. The report identifies EfficientNet-B0 as the deployment choice because it achieved the same perfect test accuracy as MobileNetV3-Large and DenseNet121 while being the smallest and fastest model among the four candidates. fileciteturn263file0L115-L132

Model	Parameters	Model Size	Test Accuracy	Throughput	Latency / Image
ResNet18	11,179,077	42.72 MB	99.78%	164.67 img/s	6.07 ms
MobileNetV3-Large	4,208,437	16.26 MB	100.00%	184.06 img/s	5.43 ms
EfficientNet-B0	4,013,953	15.61 MB	100.00%	207.06 img/s	4.83 ms
DenseNet121	6,958,981	27.14 MB	100.00%	169.80 img/s	5.89 ms

The measured efficiency results come from the project benchmark: EfficientNet-B0 had the fewest parameters, smallest model file, highest throughput, and lowest latency. Therefore, its selection was based on the overall deployment trade-off rather than accuracy alone. fileciteturn263file0L171-L181

ResNet18

ResNet18 was used as a strong baseline. It achieved 99.78% test accuracy, with one recorded Myringosclerosis image classified as Normal. fileciteturn263file0L141-L148

MobileNetV3-Large

MobileNetV3-Large is a lightweight architecture designed for efficient inference. It achieved 100.00% accuracy on all 445 test images while using only 4.21 million parameters. fileciteturn263file0L141-L151

EfficientNet-B0

EfficientNet-B0 provided the best deployment-oriented balance in the comparison. It achieved 100.00% test accuracy with 4.01 million parameters, a 15.61 MB model, and 4.83 ms measured latency per image. fileciteturn263file0L152-L154

DenseNet121

DenseNet121 also achieved 100.00% test accuracy and perfect class-wise precision, recall, and F1-score, but it was larger and slower than EfficientNet-B0. fileciteturn263file0L155-L157

---

Data Leakage Investigation

A major part of the project was identifying and correcting near-duplicate leakage.

An initial conventional stratified 70/15/15 split showed suspicious visual overlap: 185 of 439 test images (42.14%) had a pHash match in the training set at the selected threshold. A perceptual-hash grouping and union-find procedure was then used to keep visually similar images within the same split. The final leakage-safe split had zero group overlap between training, validation, and testing. fileciteturn263file0L66-L89

Leakage Audit	Result
Images analyzed	2,922
pHash grouping threshold	≤ 4
Total groups	1,924
Near-duplicate groups	509
Largest group	10 images
Cross-class near-duplicate groups	0

This is important because the final reported model results are based on the leakage-safe split rather than the original conventional random split. fileciteturn263file0L72-L89

---

Robustness Evaluation

A mild robustness dataset was generated from the leakage-safe test set using small rotations and brightness/contrast changes. EfficientNet-B0 achieved 100.00% accuracy under those recorded perturbations. This is evidence of stability for the tested transformations, but it does not establish robustness to arbitrary blur, noise, illumination changes, cropping, compression, camera differences, or other domain shifts. fileciteturn263file0L205-L212

---

Prediction Confidence Analysis

Before calibration, EfficientNet-B0 produced extremely high softmax confidence on the internal test distribution:

Statistic	Confidence
Minimum	0.9967932
Mean	0.99992484
Median	0.9999796
Maximum	1.0

No test image had confidence below 0.99. However, the project explicitly treats these values as model confidence rather than clinical certainty. External tests demonstrated that high confidence can still occur under domain shift. fileciteturn263file0L186-L199

---

External Image Sanity Checks

The selected EfficientNet-B0 model was also manually tested on independently sourced images outside the training and formal test sets. These checks were qualitative sanity tests, not a clinical benchmark. fileciteturn263file0L235-L250

Recorded examples included:

External Image	Model Output	Observation
User-labeled Acute Otitis Media	Acute Otitis Media, 99.95%	Correct qualitative result
User-labeled Myringosclerosis	Chronic Otitis Media, 47.23%	Incorrect / uncertain
Another user-labeled Myringosclerosis	Normal, 57.79%	Incorrect / uncertain
User-labeled Acute Otitis Media	Acute Otitis Media	Correct qualitative result
Chair image	Chronic Otitis Media, 72.48%	Clear out-of-distribution failure

These observations demonstrate why the internal 100% test accuracy must not be presented as universal or clinical generalization. The report specifically highlights Myringosclerosis as an area requiring additional independent data. fileciteturn263file0L243-L250

---

Experimental Workflow

The complete experimental workflow included:

1. Dataset inspection
2. Exact duplicate removal
3. Perceptual near-duplicate analysis
4. Group-aware leakage prevention
5. Consistent preprocessing and augmentation
6. Multi-model transfer learning
7. Validation-loss checkpoint selection
8. Model comparison
9. Accuracy, precision, recall, F1 and confusion-matrix evaluation
10. Model size and inference-speed benchmarking
11. Prediction confidence analysis
12. Mild perturbation robustness testing
13. Grad-CAM visualization
14. Quantitative Grad-CAM faithfulness evaluation
15. Confidence calibration
16. External-image sanity checks
17. Model packaging and web deployment

The report describes this as a complete five-class deep-learning pipeline and emphasizes the distinction between the original random split and the final leakage-safe evaluation. fileciteturn263file0L314-L327

---

Reproducibility and Saved Artifacts

The final inference package contains:

models/
├── efficientnet_b0_ear_disease.pth
├── class_mapping.json
└── temperature_scaling.json

The class mapping is:

0 → Acute Otitis Media
1 → Cerumen Impaction
2 → Chronic Otitis Media
3 → Myringosclerosis
4 → Normal

The selected checkpoint and class mapping are documented in the project report, while the current application additionally stores the fitted temperature-scaling parameters for calibrated inference. fileciteturn263file0L255-L267

---

Dataset

The project started with 3,000 original images.

After removing 78 exact duplicate copies, the leakage-safe dataset contained:

2,922 unique images

A group-aware split was used to prevent related/duplicate images from crossing dataset boundaries:

Split	Images
Training	2,059
Validation	418
Test	445
Total	2,922

Classes:

Class	Original Images
Acute Otitis Media	600
Cerumen Impaction	600
Chronic Otitis Media	600
Myringosclerosis	600
Normal	600

---

Model Training

The training pipeline uses staged transfer learning:

1. Load pretrained ImageNet EfficientNet-B0 weights.
2. Replace the final classifier with a 5-class classifier.
3. Initially freeze the backbone and train the classification head.
4. Unfreeze the upper layers and fine-tune.
5. Use AdamW optimization.
6. Use CrossEntropyLoss.
7. Use ReduceLROnPlateau scheduling.
8. Save the checkpoint with the lowest validation loss.
9. Apply early stopping.

Important training settings include:

Batch size:       32
Learning rate:    1e-4
Weight decay:     1e-4
Image size:       224 × 224
Random seed:      42

---

Evaluation

On the leakage-safe internal test set of 445 images, the final EfficientNet-B0 model achieved:

Metric	Score
Accuracy	100.00%
Balanced Accuracy	100.00%
Macro Precision	100.00%
Macro Recall	100.00%
Macro F1	100.00%
Weighted F1	100.00%

Test-set class support:

Class	Test Images
Acute Otitis Media	61
Cerumen Impaction	105
Chronic Otitis Media	68
Myringosclerosis	101
Normal	110

Important evaluation caveat

These results are from the project's internal leakage-safe test set. They should not be interpreted as clinical or real-world accuracy.

External images showed domain shift, particularly for Myringosclerosis. The system also does not include a dedicated unknown/out-of-distribution class, so arbitrary non-ear images may still receive a prediction.

---

Confidence Calibration

Temperature scaling was fitted using the validation set and then evaluated on the untouched test set.

Temperature = 0.631204

Test-set calibration results:

Metric	Before Calibration	After Calibration
NLL	0.00005482	0.00000024
ECE	0.00005478	0.00000012

Temperature scaling changes the probability distribution while preserving the model's predicted class.

---

Explainable AI — Grad-CAM

Grad-CAM is generated from the final EfficientNet feature layer to visualize image regions contributing to the prediction.

A quantitative deletion-faithfulness evaluation was also performed on all 445 test images:

Top attribution pixels masked	Predicted-class probability drop	Random-mask drop
10%	2.87 pp	0.03 pp
20%	17.48 pp	0.12 pp
30%	31.26 pp	0.78 pp
50%	50.12 pp	8.80 pp

The substantially larger probability drop for high-attribution regions compared with random masking provides evidence that the Grad-CAM explanation is faithful to the model's prediction.

Grad-CAM is an interpretability tool, not a clinically validated lesion-localization method.

---

Image Quality Checks

Before inference, uploaded images are checked for:

• JPEG/PNG format
• Valid image contents
• Valid dimensions
• Maximum pixel count
• Minimum resolution of 224 × 224
• Excessive blur
• Excessive darkness
• Excessive brightness
• Insufficient contrast

Current upload limit:

50 MB per request

---

Security

The backend includes several production-oriented protections:

• JWT authentication
• Access and refresh token support
• Token blacklist on logout
• Per-user history ownership checks
• Protected original images and Grad-CAM media
• MIME type and actual image-format validation
• Secure filenames
• Maximum request size
• Image pixel limit
• API rate limiting
• Configurable CORS origins
• Environment-based secrets and database configuration
• User-specific history deletion
• Authenticated PDF report access

For production, Redis is recommended as the rate-limit storage backend instead of the default in-memory store.

---

API

Health

GET /health

Returns API, model, database, and version status.

Model information

GET /model-info

Returns model architecture, version, input size, classes, and calibration status.

Authentication

POST /signup
POST /login
POST /refresh
POST /logout

Prediction

POST /predict

Requires JWT authentication and accepts an image upload.

History

GET    /history
DELETE /history/<history_id>
DELETE /history

Protected media

GET /media/<filename>

PDF report

GET /report/<history_id>

---

Tech Stack

Frontend

• React
• Vite
• Tailwind CSS
• Axios
• React Router
• React Dropzone
• Chart.js

Backend

• Python
• Flask
• Flask-JWT-Extended
• Flask-CORS
• Flask-Limiter
• Gunicorn

Machine Learning

• PyTorch
• Torchvision
• EfficientNet-B0
• OpenCV
• Pillow
• Grad-CAM

Database

• MongoDB
• PyMongo

Reporting

• ReportLab

Testing / CI

• pytest
• GitHub Actions

---

Project Structure

Ear_Disease_Detection/
│
├── backend/
│   ├── model/
│   │   ├── train.py
│   │   ├── evaluate.py
│   │   ├── calibrate.py
│   │   ├── evaluate_calibration.py
│   │   ├── evaluate_gradcam.py
│   │   └── analyze_image_quality.py
│   │
│   ├── routes/
│   │   ├── auth_route.py
│   │   ├── predict_route.py
│   │   ├── history_route.py
│   │   ├── report_route.py
│   │   └── info_route.py
│   │
│   ├── utils/
│   │   ├── predict.py
│   │   ├── gradcam.py
│   │   ├── image_quality.py
│   │   └── llm_agent.py
│   │
│   ├── database/
│   ├── tests/
│   ├── app.py
│   ├── config.py
│   ├── extensions.py
│   ├── requirements.txt
│   └── requirements-dev.txt
│
├── frontend/
│   ├── src/
│   ├── package.json
│   └── ...
│
├── models/
│   ├── efficientnet_b0_ear_disease.pth
│   ├── class_mapping.json
│   └── temperature_scaling.json
│
├── scripts/
│   └── prepare_dataset.py
│
├── evaluation/
│   └── efficientnet_b0/
│
└── .github/
    └── workflows/
        └── ci.yml

---

Local Setup

1. Clone

git clone https://github.com/SRV30/Ear_Disease_Detection.git
cd Ear_Disease_Detection

2. Backend

Create and activate a virtual environment:

Windows PowerShell

python -m venv venv
venv\Scripts\activate

Install dependencies:

python -m pip install -r backend/requirements.txt
python -m pip install -r backend/requirements-dev.txt

Create backend/.env:

MONGO_URI=mongodb://localhost:27017
JWT_SECRET_KEY=replace-with-a-long-random-secret
CORS_ORIGINS=http://localhost:5173,http://127.0.0.1:5173
RATE_LIMIT_STORAGE_URI=memory://

Start the backend:

python backend/app.py

The API runs on:

http://127.0.0.1:5000

3. Frontend

cd frontend
npm install

Create frontend/.env:

VITE_API_BASE_URL=http://127.0.0.1:5000

Start the development server:

npm run dev

The frontend is normally available at:

http://localhost:5173

---

Testing

Run the backend automated tests:

pytest backend/tests -v

The test suite covers:

• API health
• Model metadata
• Authentication protection
• Protected history/media/report routes
• Image validation
• Invalid image rejection
• MIME/extension validation
• Oversized image rejection

The project also has a GitHub Actions workflow that runs backend tests and the frontend production build on pushes to main and pull requests targeting main.

---

Environment Variables

Backend

MONGO_URI=
JWT_SECRET_KEY=
CORS_ORIGINS=
RATE_LIMIT_STORAGE_URI=

Frontend

VITE_API_BASE_URL=

Do not commit real secrets or production credentials.

---

Screenshots

Project screenshots/demo material:

https://drive.google.com/file/d/1Sfiavxmrg6wtNisTqRmeEYrrCt03_czZ/view?usp=sharing

---

Limitations

• Internal test performance does not establish clinical performance.
• Real-world domain shift can reduce reliability.
• The model can still produce predictions for images outside the five target classes.
• Image quality and acquisition conditions affect predictions.
• Grad-CAM explanations are not clinical lesion localization.
• Further external validation on diverse clinical datasets is required.
• The system should be used as decision support, not as a replacement for a healthcare professional.

---

Future Work

• Larger and more diverse clinical datasets
• External validation across different acquisition devices and clinical sites
• Improved handling of unknown/out-of-distribution images
• Mobile deployment
• Hospital/EHR integration
• Additional model architectures and ensemble methods
• More comprehensive clinical validation

---

Author

SRV

---

License

This project is intended for academic and research purposes.

---

Acknowledgements

• Medical imaging and otoscopy research community
• PyTorch and Torchvision
• Flask ecosystem
• React ecosystem
• Open-source explainable AI research