eduzhai > Applied Sciences > Engineering >

Name that manufacturer Relating image acquisition bias with task complexity when training deep learning models experiments on head CT

  • king
  • (0) Download
  • 20210506
  • Save

... pages left unread,continue reading

Document pages: 15 pages

Abstract: As interest in applying machine learning techniques for medical imagescontinues to grow at a rapid pace, models are starting to be developed anddeployed for clinical applications. In the clinical AI model developmentlifecycle (described by Lu et al. [1]), a crucial phase for machine learningscientists and clinicians is the proper design and collection of the datacohort. The ability to recognize various forms of biases and distributionshifts in the dataset is critical at this step. While it remains difficult toaccount for all potential sources of bias, techniques can be developed toidentify specific types of bias in order to mitigate their impact. In this workwe analyze how the distribution of scanner manufacturers in a dataset cancontribute to the overall bias of deep learning models. We evaluateconvolutional neural networks (CNN) for both classification and segmentationtasks, specifically two state-of-the-art models: ResNet [2] for classificationand U-Net [3] for segmentation. We demonstrate that CNNs can learn todistinguish the imaging scanner manufacturer and that this bias cansubstantially impact model performance for both classification and segmentationtasks. By creating an original synthesis dataset of brain data mimicking thepresence of more or less subtle lesions we also show that this bias is relatedto the difficulty of the task. Recognition of such bias is critical to developrobust, generalizable models that will be crucial for clinical applications inreal-world data distributions.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×