Automated Pneumonia Detection from Chest X-Rays: Overcoming Data Imbalance with VGG16

In diagnosing and screening for pneumonia, extensive research is underway to avoid deliberate delays caused by the shortage of human radiologists. Here, we implemented transfer learning on randomly generated image data to develop a robust diagnostic tool. The construction of the CNN starts with randomly sampled images of healthy and infected lungs, which are then augmented and weighted by class. Using this pursuer, we drafted the CNN in two variants: one without dropout noise (Baseline-CNN) and the other with noise to mitigate overfitting attacks (VGG16-Transfer). The randomness applied to the training batches provides significant defence against false-positive attacks. This results in predictions that are largely unrecognisable as basic errors and, on the other hand, are easy for human doctors to trust, yielding comparatively better reliability.