top of page

2D Medical Imaging - 18x faster

Dataset: MICCAI Gleason 2019 Challenge Dataset - https://gleason2019.grand-challenge.org/

MICCAI 2019 Automatic Prostate Gleason Grading Challenge: This challenge aims at the automatic Gleason grading of prostate cancer from H&E-stained histopathology images. This task is of critical importance because Gleason score is a strong prognostic predictor. On the other hand, it is very challenging because of the large degree of heterogeneity in the cellular and glandular patterns associated with each Gleason grade, leading to significant inter-observer variability, even among expert pathologists.

Machine config: 8 x NVIDIA A100-SXM4-40GB


Machine config: 5 x A6000


Model architecture: UNet2D


Remarks w/ Throughput:


  • With DLOP:

    • Configuration: {batch_size: 128, prefetch_factor: 8, num_workers: 32}

    • Average throughput: 810 imgs/sec


  • Without DLOP:

    • Configuration: {batch_size: 128, prefetch_factor: 2, num_workers: 64}

    • Average throughput: 83 samples/sec




Speedup: ~10x over regular training.

Comments


bottom of page