Hyperparameter optimization in transfer learning for medical image analysis
Rune Johan Borgli
Medical images from gastroenterology
Gastrointestinal (GI) tract prone to many different diseases
Doctors use sensor and image data to diagnose
The procedure's success heavly dependent on doctor
Images from colonoscopies of different conditions
The problem
We use standard Bayesian optimization
Difficult to create large datasets required for deep learning
Transfer learning is a solution
Choosing hyperparameters for fine-tuning is unintuitive
Choosing delimiting layer in fine-tuning is complex and previous work has done this manually
We optimize with regards to validation accuracy
Three ways to structure the optimization
The hyperparameters we optimize are the pre-trained model, optimizer, learning rate and delimiting layer
our approach
Complex structure and very difficult to know which layers to tune
Results from 8-class dataset
Separate optimizations
Separate hyperparameters
Shared hyperparameters
Layer optimization
Validation accuracy
Validation accuracy
Epoch
Time
Model
InceptionResNetV2
Optimizer
Adadelta
Learning Rate
0.84
Layer
136
Conclusion
Automatic hyperparameter optimization with our hyperparameters including delimiting layer is successful in achieving better results
The delimiting layer from the automatic hyperparameter optimization is not trivial to choose manually
Shared hyperparameters gave the best results