AI Otoscope Smartphone Ear Infection

ISEF Category: Biomedical Engineering

Subcategory: Biomedical Devices  ·  Difficulty: Advanced  ·  Setup: Home Setup  ·  Time: 1 to 2 Months

The Hook

Ear infections send kids to the doctor more than almost any other complaint, yet diagnosis still depends on a tiny handheld scope and a tired clinician. A $10 endoscope camera that clips into a smartphone, plus a YOLO model trained on public ear images, gets close to expert performance. Active learning even helps it improve on every new patient.

What Is It?

An otoscope is a magnifying lens with a light used to see the eardrum. Replacing it with a USB endoscope camera lets a smartphone view the same scene at higher resolution.

YOLO is a real-time object-detection model. Trained on the Hadassah and Chile otoscopy datasets, it draws bounding boxes around eardrum regions and classifies them as healthy or infected.

On-device active learning lets the model flag uncertain images for a human label and update over time. This keeps the user privacy-safe while improving performance.

Why This Is a Good Topic

Pediatric ear care is a real screening problem. Hardware is cheap, datasets are open, and modeling is well-trodden. You will learn data labeling, object detection, and active-learning evaluation.

Research Questions

• How does training-set size change YOLO balanced accuracy?
• What is the effect of image resolution on detection rate?
• Does active learning reduce required labels to reach a target accuracy?
• To what extent does device-side inference latency vary by phone model?
• Which preprocessing (color normalization) boosts cross-dataset transfer?
• How does label noise from raters affect outcomes?
• What is the effect of glare and wax artifacts on misclassification?

Basic Materials

• USB endoscope ear camera ($10 range).
• Smartphone with OTG adapter.
• 3D-printed otoscope clip.
• Public Hadassah or Chile otoscopy dataset (with permissions).
• Informed-consent form for any live use.

Advanced Materials

• Clinical otoscope for ground-truth comparison.
• Cloud GPU.
• Pediatric mentor.
• IRB approval if collecting new data.

Software & Tools

• PyTorch and Ultralytics YOLO: Trains the detection model.
• Roboflow: Manages dataset labeling.
• TensorFlow Lite: Deploys to mobile.
• Python (scikit-learn): Computes ROC and calibration.

Experiment Steps

1. Lock the camera clip geometry.
2. Build a strict subject-wise dataset split.
3. Decide label classes and preprocessing.
4. Train YOLO and validate on held-out subjects.
5. Implement an active-learning loop and measure label efficiency.
6. Run fairness slices and report balanced accuracy per slice.

Common Pitfalls

• Treating the project as diagnostic instead of screening.
• Mixing datasets without color normalization.
• Reporting only mAP without confidence intervals.
• Letting active learning sample only easy cases.
• Storing user images without consent.

What Makes This Competitive

A competitive project uses strict subject-wise splits, reports balanced accuracy and confidence intervals, and runs fairness slices across skin and age groups in the datasets. Active-learning evaluation needs a controlled study comparing labels-needed vs. accuracy gained.

Project Variations

• Compare YOLO with a vision transformer baseline.
• Add a tympanic membrane mobility test using a puff of air.
• Build a multilingual app interface for global accessibility.

Learn More

• PubMed: Search smartphone otoscope deep learning reviews.
• NIH PubMed Central: Open-access pediatric otitis media papers.
• Ultralytics YOLO documentation: Free training guides.
• Roboflow Universe: Curated otoscopy datasets.
• MIT OpenCourseWare: Course 6.S191 Introduction to Deep Learning.