Multi classfication of medical images

Part 1 (2019)
1

+Newbie alert

I followed the steps mentioned in lesson 3 https://github.com/fastai/course-v3/blob/master/nbs/dl1/lesson3-planet.ipynb on fundus (medical) image dataset (400 images) . I tried two different image augmentations
get_transforms(flip_vert=True, max_lighting=0.1, max_zoom=1.05, max_warp=0.) // as in lesson 3

and
tfms = get_transforms(max_rotate=30, max_zoom=1.3, max_lighting=0.4, max_warp=0.0,
p_affine=1., p_lighting=1.)

Even after multiple levels , lr find -> train-> unfreeze-> lr find -> train , followed by two sets of training on higher resolution images , I am not able to cross 83-84 % accuracy .
I tried on both densenet101 and resnet101 pretrained models . Didnt get good results in resnet50

As I understand , dataset is much from ordinary datasets . My gut feeling says I need to train many cycles to get closer to 92-93 %
Any other suggestions on how to improve accuracy ?

Thanks in advance

2

What type of medical image (microscopic/DICOM/WSI)? Please share your kernel for others to look and comment at.

3

Images look something like https://images.app.goo.gl/LQfiPjuAEUNQ6B186 , its not dicom , its a normal tiff image.

4

You can look at the approaches here

1 Like
5

Can you share your notebook.
It would be helpful.