Convolutional neural network based on the skin lesion image data set HAM10000.
Created as part of Project 2 for module COS624 (Cognitive Systems M.Sc.).
Results
I trained 5 versions of the ConvNet, using 5 different data sets:
- Data set with 8x8 pixel images (as provided in the HAM data set).
- Data set with 28x28 images (as provided in the HAM data set).
- Data set with 28x28 images, same as above, but with augmented lesion types to 2000 images per category.
- Data set with 42x42 images, that I created, with augmented lesion types to 2000 per category.
- Data set with 64x64 images, that I created, with augmented lesion types to 2000 per category.
All networks were trained for 30 epochs. The table below summarises the results:
| Test accuracy | Test loss | Training time | Epochs | |
|---|---|---|---|---|
| 8x8 (as provided) | 0.710934 | 0.86682 | 0:00:49 | 30 |
| 28x28 (as provided) | 0.729905 | 0.785197 | 0:06:07 | 30 |
| 28x28 (augmented) | 0.62069 | 0.999897 | 0:11:25 | 30 |
| 42x42 (augmented) | 0.659984 | 0.944596 | 0:25:51 | 30 |
| 64x64 (augmented) | 0.652767 | 1.10109 | 0:54:19 | 30 |
Additionally, the figure below shows the training histories, outcomes and confusion matrices for all networks:
