Disease classification with educational medical images
An educational and research project only: its results are not a medical diagnosis and are not used to make any treatment decision.
Problem description
Building an educational image classifier on openly licensed medical images (such as the small MedMNIST datasets, most of them under CC BY 4.0 except DermaMNIST, which is under the non-commercial CC BY-NC 4.0) using transfer learning in Teachable Machine or Image Embedding in Orange. An educational and research project only: its results are not a medical diagnosis and must not be used for a treatment decision.
Objectives
- Understanding the privacy and ethics of medical images
- Applying transfer learning to medical images
- Evaluating the model with Recall and the Confusion Matrix
Problem type
Image Classification
Tools used
Algorithms
Implementation steps
- Open medical-image-classification.ows in Orange after installing the Image Analytics extension (Options → Add-ons), then choose your image folder in Import Images (it uses the local SqueezeNet embedder, so the images are not uploaded).
- Choose an openly licensed educational medical image dataset from its official source and read its terms (do not use real patient images without consent).
- Teachable Machine: Image Project → a class for each condition → upload the images → Train → test with images that were not in the training.
- Orange (Image Analytics extension): Import Images → Image Embedding → Test and Score with Logistic Regression.
- Document the source and licence of the images in your report.
Workflow diagram
Orange
Image AnalyticsEvaluateModel
↔ Drag the diagram sideways to see all of it.
Connections as text (5)
- Import Images → Image Embedding (Data → Images)
- Image Embedding → Test and Score (Embeddings → Data)
- Logistic Regression → Test and Score (Learner)
- Test and Score → Confusion Matrix (Evaluation Results)
- Import Images → Image Viewer (Data)
Project files
Results
No published results
Evaluation metrics
- Recall for each class
- Confusion Matrix
- ROC AUC for binary cases
Interpreting the results
Transfer learning usually gives good results with few images, but performance on images from a different hospital or device may drop a lot (Domain Shift). An educational and research project only: its results are not a medical diagnosis and must not be used for a treatment decision.
Discussion questions
- Why might the model fail on images from another X-ray machine?
- What approvals are required to use real patient images?
- How would you explain the model's decision to a doctor (XAI)?
Try it in the Visual Neural Network Lab
