Having written hand-crafted networks for years, I have recently been testing the meta-experiments framework: you conduct thousands of AI experiments based on a range hyper-parameter values, do ablations, and choose the best performer. I have a near-mature repo which joins the strengths of ray-tune (generates random AI trials, multi-GPU support), neptune (tracks trial progress,metrics, and charts), and fastai. I am seeking interested individuals for a collab in radiology/cancer AI on 3D datasets (#KiTS, #LiTS challenge). Screenshots below showcase the features:
Neptune in-depth: charts from a single experiment. Do you spot how the LR (figure on top-left) is flat and later decays. That’s the ReduceLROnPlateau callback!
A callback uploads training / validation results per epoch onto neptune in a color-coded scheme(3-classes 3 colors for KiTS). The image below is not captioned so, from the left images are in threes, a greyscale input image, color-coded prediction, (red=kidney, green=tumour) and color-coded ground truth. You will see a lot of black images, because they are slices without tumour.
Caveat:
This is a highly opinionated quasi-library of tools, designed with a very specific task in mind: medical image segmentation and classification.
P.S I work mostly with neovim highly adapted to fast python coding. Interested individuals may contact to share configs.
Dear Dr. Usman,
I am very interested to work on this project. However, I am a medical student and still a beginner at deep learning, I wrote some codes and have experience with AI in medicine. Also, I can be of benefit to you in the clinical part I am a clinical researcher with 25 publications and an active biostatistician with 2 years of experience with R and machine learning in R and Python.
Best Regards
Thanks for getting in touch. Please stay posted for updates and when I upload the project. It will be free to discuss and contribute at that time. In the meanwhile, I have sent you a DM to discuss options for a collab while the project is in its current stage.
Hi. Not really. I have considered using them for data augmentation, and I might add them in. The way its setup, the library currently supports UNet flavours you can customise. I have just copy-pasted nnUNet’s model that can be tried. I am getting very high scores for kidney (DC .95) and tumour (DC .86) based on this scheme.
