Detection of pneumonia on X-Rays

Binary classification

Для нахождения пневмонии была обучена нейросеть ResNet50 на фреймворке PyTorch

Firstly, I have attempted to train ResNet152 CNN without weights, using Tensorflow. I've obtained accuracy = 0.92 on the train dataset, accuracy = 0.89 on the validation dataset. Recall - 0.88. It meaned overfitting.

Further, I've decided to apply another approach and used PyTorch. I chose ResNet50 with weights, freezed all the layers and added linear classifier (with 2 outputs). It took only 10 epochs to get enough good results.

Results achieved: train acc = 0.94, validation acc = 0.97, test acc = 0.95. Precision = 0.97, Recall = 0.97.

Thus, it turned out to increase Recall from 0.88 to 0.97 using transfer learning. FalseNegatives have been reduced from 44 to 25.

Used image dataset from Kaggle: Chest X-Ray Images (Pneumonia). I have applied the only one augmentation: RandomHorizontalFlip.

I have prepared train, val and test datasets in Roboflow www.roboflow.com .

Some test images:

Confusion matrix:

This is my train pet-project.