Added image drawing and imported the drawn image to the neral network. So far the predictions are correct 40% of the time, I suspect thats because of the sharper use drawing mode compared to my training images which are rounder and more blended in yo the background. It also took me a while to realise the MNIST images have black backgrounds