(Project during AI for Healhcare - Udacity)
Quickly and effectively identify positive pneumonia cases from radiological images (DICOM) using a machine learning model.
Data Preprocessing: Verify DICOM data and generate ground truth for 14 common thoracic pathologies using NLP. Data Augmentation: Normalize and split data into validation and training sets. Model Building: Use a pre-trained model, adjust parameters, and build a custom model. Training and Optimization: Compile the model, train with callbacks, and save the best weights. Validation and Deployment: Evaluate performance, make predictions, and save the final model.
Classification Threshold: Set at 0.40 to maximize true positives. Performance Metrics:
Accuracy: 0.452 Precision: 0.310 Recall: 0.692 F1 Score: 0.429
The model prioritizes high recall to capture as many positive pneumonia cases as possible, sacrificing precision.
This project emphasizes detecting pneumonia cases with a focus on minimizing false negatives.
See the report for more details.