This Repo contains the classification model built on MR images for the classification of brain tumor.
- The data contains 4 classes:
- No Tumor
- Pituitary
- Glioma
- Meningioma
- The datset can be found here.
- The
no_tumor
class contains half the examples than the other 3 classes. To mitigate this problem, the data from Br35H is used to increase the examples in theno_tumor
class and its size is made equal to the other 3 classes.
- The architecture used for the classification of is Pre-trained EfficientNet-B0.
- Optimizer:
Adam
- Loss:
Categorical CrossEntropy
The model is evaluated on 394 test images, unseen by the trained model before.
- Version 1:
- The training and validation accuracy came out to be
~95%
. - The test accuracy proved to be
~71%
yet. - The worst learned class being
no_tumor
, probably because of class imbalance in the training data whereno_tumor
is almost half the amount of other classes in the original data.
- The training and validation accuracy came out to be
- Version 2
- Added gaussian Noise in training data.
- The training and validation accuracy came out to be
~97%
. - The test accuracy proved to be
~76%
.