Tuberculosis

The purpose of this project is to check how well a computer vision model built from scratch performs against pre- trained model (VGG16, VGG19, ResNet, InceptionV3 …). Datasets are X-ray images of patient chests with tuberculosis. Interestingly, the model built from scratch performed very well, achieving an accuracy of 97% on the validation dataset and an accuracy of 95% on the test dataset. The best score was achieved by VGG19 (more than 99% on the validation dataset and test dataset) after using transfer learning techniques – fine tuning.

Two methods are mainly used here in order to achieve our goal. Transfer Learning techniques (Feature Extraction && Fine Tuning) and the other is about building a classfier model from scratch (CNN from scratch).

The dataset can be downloaded from kaggle: https://www.kaggle.com/tawsifurrahman/tuberculosis-tb-chest-xray-dataset

Result of the experiment (Alphabetical Order)

CNN from scratch:

- Validation accuracy: 97.35 %
- Validation loss: 8.34 %

- Test accuracy: 95.63 %
- Test loss: 12.82 %

DenseNet201:

- Validation accuracy: 98.12 %
- Validation loss: 5.20 %

- Test accuracy: 97.50 %
- Test loss: 7.21 %

EfficientNetB3:

- Validation accuracy: 93.82 %
- Validation loss: 18.21 %

- Test accuracy: 95.63
- Test loss: 15.45 %

InceptionV3 - Fine Tuned:

-Validation accuracy: 97.06 %
-Validation loss: 7.16 %

-Test accuracy: 98.75 %
-Test loss : 4.83 %

ResNet50V2:

- Validation accuracy: 97.94 %
- Validation loss: 5.04 %

- Test accuracy: 98.12 %
- Test loss : 5.58 %

VGG16 - Fine Tuned:

- Validation Accuracy: 99.37 %
- Validation Loss: 0.85 %

- Test Accuracy: 99.37 %
- Test Loss: 0.99 %

VGG19 - Fine Tuned:

- Validation accuracy: 99.85 %
- Validation loss: 1.71 %

- Test accuracy: 99.37 %
- Test loss: 3.30 %