Concept Guide

6 Addressing the Memory Bottleneck in AI Model – Training for Healthcare
• Image size: Images are often down sampled to a lower resolution
• Batch size: Batch sizes are often reduced to one or two images
• Tiling/Patching: Images are often subsampled into overlapping tiles/patches
• Model Complexity: Reductions in the number of feature maps and/or layers are often
necessary
• Model Parallelism: Models may be distributed across several compute nodes in a parallel
fashion
Although these tricks have been used to produce clinically-relevant models, we believe that
researchers would not choose to use them if it were not for the memory limitations in hardware.
In other words, these tricks were not created to obtain better models—they are instead necessary
workarounds for hardware limitations. We believe that researchers would prefer to use the full
resolution image without having to account for hyperparameters such as batch size, model
complexity, or subsampling (tiling/patching). The large memory capacity of the 2
nd
Generation
Intel Xeon Scalable Processor allows researchers this ability.
Experimental Data
The medical decathlon dataset [4] is a 3D semantic segmentation challenge with a broad range
of medical imaging tasks including tumor and cancer diagnoses for various parts of the human
body, including the liver, brain, lung, colon, and prostate. The images were generated either
through a CT or an MRI scan at various universities and research centers from across the globe.
Given this variety of data, the images present the opportunity for data scientists and machine
learning practitioners to optimize AI algorithms for generalizability in medical imaging tasks with
a primary focus on semantic segmentation. Thus, the most commonly used metric in
segmentation tasks, Dice Similarity Coefficient (DSC) [5], along with Normalized Surface Distance
(NSD) (distance between reconstructed surfaces) are used to assess different aspects of the
performance of each task and region of interest. In this paper, we focus on the DSC (or simply,
“dice coefficient”) of the Brain Tumor task from the BraTS dataset, which contains 750 4D MRI
volumes: 484 for training and 266 for testing.