Pneumonia Detection in Chest X-Rays using Transfer Learning and TPUs

Abstract

Pneumonia is a severe respiratory disease with potentially life-threatening consequences if not promptly diagnosed and treated. Chest X-rays are commonly employed for pneumonia detection, but interpreting the images can pose challenges. This study explores the efficacy of four popular transfer learning models, namely VGG16, ResNet, InceptionNet, and DenseNet, alongside a custom CNN model for this task. The model performance is evaluated using Mean Absolute Error (MAE) as the performance metric. The findings reveal that VGG16 outperforms the other transfer learning models, achieving the lowest MAE (66.19). To optimize the model training process, a distributed training strategy utilizing TensorFlow's TPU (Tensor Processing Unit) strategy is implemented. The custom CNN model is parallelized using TPU's multiple instances available over the cloud, enabling efficient computation parallelization and significantly reducing model training times. The experimental results demonstrate a remarkable decrease of 68.36% and 54.74% in model training times for the CNN model when trained using TPU compared to training on a CPU and GPU, respectively.