Is it necessary to know the pixel size of your training dataset before CNN training?

I am analyzing some medical image data from the bone age algorithm competition and when i tried to run some image .shape i noticed they had some variance


so you can see i have images that have

1653 x 1334 pixel
2066 x 1650 pixel
2043 x 1668 pixel

Is the image pixel size important or only necessary when you will do transformations ?

What would be a good value to set my ‘sz’ variable – i want to use resnet50 and i know the images trained were 256 x 256 pixel images

Is there a better model to use ? I am approaching this problem as

  1. Test using the pretrained model
  2. Since i know the ground truth for this dataset - use that to train the model
1 Like

Starting with 256 sounds fine. See how it goes. I don’t know what resolution you’d need to be able to see the medical features - try looking yourself at a 256x256 image and see if it’s still usable for you.