mongoslabs

Dataloader that serves MRI images from a mongodb.

The main idea is to keep MRI images and corresponding training labels for segmentation tasks in a mongo database. However, each 256³ 3D MRI tensor even in 8 bit precision is 16Mb. mongo's records cannot be larger than this limit and we need to also store the labels of the same dimensions. mongoslabs fetches and aggregates each 256³ tensor stored across 8 records, together with corresponding labels either for gray and white matter, 104 regions atlas, or a 50 region atlas. (The scripts to populate such collection are upcoming.)

An example of maintaining a high utilization on 4 GPUs

installation

pip install mongoslabs

usage

A detailed example of how to create a dataloader using provided dataset class and the corresponding tools is in scripts/usage_example.py

Do not forget to move the batches to the GPU once obtained.