1. Summary
What to include: A concise description of what you will build, test, and learn.
This project explores how artificial intelligence and machine learning can be used to classify images of skin lesions. Using Google Teachable Machine, an image classification model will be trained with labeled examples of benign and malignant skin lesions.
The project investigates how training image quantity, image quality, and dataset diversity influence classification accuracy. The model will be tested using previously unseen images to determine how reliably it distinguishes the selected categories. The goal is to demonstrate the potential and limitations of artificial intelligence in medical image analysis, not to develop a clinically validated diagnostic tool.
2. Engineering Connection
What to include: Explain how engineering and computer science relate to the problem.
Biomedical engineers, computer scientists, and medical researchers develop artificial intelligence systems to assist healthcare professionals in analyzing medical images. Machine learning algorithms can identify visual patterns in image datasets. Researchers investigate how these systems might support recognition of suspicious lesions and improve medical image analysis.
This project uses image classification, supervised learning, training datasets, and performance evaluation to demonstrate an engineering design process used in medical AI research.
3. Learning Objectives
After completing this project, I should be able to:
- Explain supervised machine learning and image classification.
- Describe differences between benign and malignant skin lesions.
- Identify appropriate medical image datasets.
- Build an image classification model using Google Teachable Machine.
- Train a model using labeled medical images.
- Measure classification accuracy on unseen test images.
- Explain false positives, false negatives, and model bias.
- Evaluate limitations of AI in medical screening.
4. Educational Standards
What to include: Relevant computer science, engineering, mathematics, and scientific research connections.
- Computer Science: Artificial intelligence, supervised learning, algorithms, data representation, model evaluation.
- Engineering: Define a problem, develop a solution, test performance, improve design.
- Mathematics: Percentages, statistical analysis, accuracy, sensitivity, specificity.
- Scientific Research: Hypotheses, controlled experiments, evidence-based conclusions.
- Ethics: Medical data privacy, fairness, reliability, responsible AI use.
Add official standards codes if your course or project submission requires them.
5. Materials List
| Material | Purpose |
|---|---|
| Computer and internet access | Develop and test the project |
| Google Teachable Machine | Train an image classifier |
| ISIC Skin Image Archive | Obtain research images with appropriate metadata and permissions |
| Google Sheets or Excel | Record predictions and calculate test results |
| Labeled lesion images | Training and evaluation |
| Research notebook | Document procedures and trials |
| Google Docs or Microsoft Word | Write final research report |
For an initial experiment, collect approximately 100–200 labeled images per class, if suitable data is available, and reserve additional images for testing. This is a classroom target, not a medically sufficient dataset.
6. Worksheets and Attachments
- Dataset inventory with image sources, diagnostic labels, and class counts
- Training log with model settings and trials
- Test spreadsheet with actual and predicted labels
- Confusion matrix
- Google Teachable Machine screenshots
- Final report with results, analysis, and conclusions
7. Pre-Requisite Knowledge
Basic computer skills, file management, and familiarity with image classification. Concepts to understand include supervised learning, labeled training data, holdout testing, and accuracy. Basic knowledge of benign and malignant skin lesions is helpful. Advanced programming is not required to create the model.
8. Introduction / Motivation
What to include: Explain why the topic matters and why you chose it.
Skin cancer is an important healthcare concern, and early recognition of suspicious lesions can influence treatment outcomes. Advances in artificial intelligence have created opportunities for researchers to study how computers identify patterns within medical images.
I selected this project to explore whether an accessible machine learning tool, Google Teachable Machine, can learn visual differences between benign and malignant lesion images. My research will investigate how accurately the model classifies previously unseen images and whether increasing the quantity and diversity of training images improves performance. The project will also examine incorrect predictions, particularly when malignant lesions are classified as benign.
9. Procedure
Part 1. Background Information
Research artificial intelligence, supervised machine learning, image classification, neural networks, transfer learning, lesion categories, training/test datasets, and false positives and negatives. Explain that a model learns statistical patterns in labeled images; it does not make a dermatologist's medical judgment.
Part 2. Before the Activity
- Open Google Teachable Machine.
- Select Get Started → Image Project → Standard Image Model.
- Choose a suitable labeled image dataset.
- Verify each image has appropriate diagnosis metadata.
- Organize training and test images into separate folders.
- Record the number of images in each class.
Part 3. Train the Model
- Set up classes: Start with “Benign Skin Lesions” and “Malignant Skin Lesions.” Do not assume a random skin photograph is benign.
- Upload images: Place verified images in the corresponding classes; document sources and keep the initial class counts reasonably balanced.
- Train: Select Train Model. Record epochs, batch size, learning rate, and other settings available. Keep settings constant in the first comparison.
- Test: Use new, unseen images. Record the predicted label, output confidence score, and known diagnosis label for each image.
Part 4. Engineering Design Challenge
Experiment: Does increasing the number of training images improve classification accuracy?
| Experiment | Benign training images | Malignant training images |
|---|---|---|
| Model A | 50 | 50 |
| Model B | 100 | 100 |
| Model C | 200 | 200 |
| Model D | 400 | 400 |
These numbers are proposed experimental targets. Use the same held-out test set for all models and hold other settings constant. Repeat training when practical and note variability between runs.
Part 5. Connect to Medical Research
- Did increasing training data improve accuracy?
- Which category was more often misclassified?
- Were incorrect predictions ever given high confidence?
- Could performance differ by skin tone or imaging method?
- Why are false negatives particularly concerning?
- Why is independent clinical validation required before medical use?
10. Vocabulary
| Term | Definition |
|---|---|
| Artificial intelligence | Technology developed to perform tasks associated with human intelligence |
| Machine learning | Methods through which computers learn patterns from data |
| Supervised learning | Training on examples with known labels |
| Dataset | Collection of examples used for research or model development |
| Neural network | Computational model that learns patterns from data |
| Classification | Assigning an input to a category |
| Benign | Not cancerous |
| Malignant | Cancerous |
| Accuracy | Fraction of all predictions that are correct |
| Sensitivity | Fraction of actual malignant cases correctly identified |
| Specificity | Fraction of actual benign cases correctly identified |
| False positive | Benign lesion predicted as malignant |
| False negative | Malignant lesion predicted as benign |
| Overfitting | Learning training-specific patterns that do not generalize well to new data |
11. Assessment
Pre-Activity Assessment
Explain your existing knowledge of AI, machine learning, and skin lesions.
Formative Assessment
- Record image counts and sources.
- Record training configurations.
- Document problems and changes between trials.
- Review whether the test dataset remained unseen during training.
Post-Activity (Summative) Assessment
Enter actual measurements after each experiment; do not replace missing measurements with estimates.
| Experiment | Test images | Accuracy | Malignant sensitivity |
|---|---|---|---|
| Model A | 100 (proposed) | To be measured | To be measured |
| Model B | 100 (proposed) | To be measured | To be measured |
| Model C | 100 (proposed) | To be measured | To be measured |
| Model D | 100 (proposed) | To be measured | To be measured |
A 100-image holdout set is an initial classroom suggestion. Report the actual malignant and benign test counts separately.
Confusion Matrix
| Actual class | Predicted malignant | Predicted benign |
|---|---|---|
| Malignant | True positive (TP) | False negative (FN) |
| Benign | False positive (FP) | True negative (TN) |
Interpret the results by discussing whether the hypothesis was supported, which mistakes occurred, and how reliable the experimental findings are.
12. Safety
Use properly licensed research images. Avoid identifiable patient photos without appropriate permission and consent. Discuss image-source bias and representation of different skin tones. A model's displayed confidence score is not the clinical probability that someone has cancer.
13. Troubleshooting
| Problem | Possible solution |
|---|---|
| Incorrect predictions | Review labels, image sources, and dataset quality |
| Good training results but poor test results | Investigate overfitting and data leakage |
| Bias toward one class | Review class balance and image quality |
| Model recognizes backgrounds | Standardize preparation and analyze source bias |
| Training takes too long | Begin with a smaller dataset |
| Results vary between runs | Repeat experiments and report variability |
| Poor subgroup performance | Review dataset diversity and report subgroup limitations |
14. References
- Google Teachable Machine — Model development platform.
- International Skin Imaging Collaboration (ISIC) — Research image archive.
- ISIC Challenge Datasets — Datasets and citation resources.
- National Cancer Institute — Skin Cancer — Medical background.
In the final report, add each specific dataset version, research paper, image license, and access date actually used.
15. Copyright & Attributions
Credit Google Teachable Machine as the model development platform and the ISIC image contributors and corresponding publications according to dataset license terms. Acknowledge that the instructional outline was adapted from Teach Engineering.org recommended outline.
16. Project Information / Quick Look
| Grade Level | High school (9–12) |
|---|---|
| Subjects | Computer Science, Biology, AI, Biomedical Engineering |
| Project Type | Experimental Machine Learning Research |
| Software | Google Teachable Machine |
| Dataset | ISIC Skin Image Archive (selected verified subset) |
| Independent Variable | Number of training images |
| Dependent Variables | Accuracy, sensitivity, specificity |
| Deliverables | Model, testing spreadsheet, figures, and final report |
| Estimated Duration | 2–3 weeks, depending on data preparation and trials |
Final Research Report Checklist
Use this checklist when drafting your completed report. Boxes can be checked in the browser; print the page to retain a paper copy.