I suppose there is an important bias in the public dataset (https://data.mendeley.com/datasets/rscbjbr9sj/3) that was used in this work. If one checks the image size distribution you see the sizes are completely different between normal and pneumonia cases. Unfortunately, information leakage occured when selecting and preparing the images for analysis. This information leakage from image size seems to propagate to other image features which overfits any AI model.