Unsupervised Data Drift Detection Using Convolutional Autoencoders: A Breast Cancer Imaging Scenario
Authors:
Date: 30.11.-0001
Abstract
Imaging AI models are starting to reach real clinical settings, where model drift can happen due to diverse factors. That is why model monitoringmust be set up in order to preventmodel degradation over time. In this context, we test and propose a data drift detection solution based on unsupervised deep learning for a breast cancer imaging setting. A convolutional autoencoder is trained on a baseline set of expected images and controlled drifts are introduced in the data in order to test if a set of metrics extracted from the reconstructions and the latent space are able to distinguish them.We prove that this is a valid tool that manages to detect subtle
differences even within these complex kind of images.
BIB_text
title = {Unsupervised Data Drift Detection Using Convolutional Autoencoders: A Breast Cancer Imaging Scenario},
pages = {345-354},
keywds = {
data drift detection · deep learning · convolutional autoencoder · medical imaging · breast cancer
}
abstract = {
Imaging AI models are starting to reach real clinical settings, where model drift can happen due to diverse factors. That is why model monitoringmust be set up in order to preventmodel degradation over time. In this context, we test and propose a data drift detection solution based on unsupervised deep learning for a breast cancer imaging setting. A convolutional autoencoder is trained on a baseline set of expected images and controlled drifts are introduced in the data in order to test if a set of metrics extracted from the reconstructions and the latent space are able to distinguish them.We prove that this is a valid tool that manages to detect subtle
differences even within these complex kind of images.
}
isbn = {978-981-99-3311-2},
date = {0000-00-00},
}
