In the Autoencoders to detect manifestation shift in medical images research, we implemented a simple autoencoder architecture, with latent space containing a 6 times image reduction of original medical images.
The key component of this research if available at autoencoder.py file.
Key elements of this architecture:
- Style transfer concepts
- Instance Normalization (IN) in shallow layers
- Batch Normalization (BN) in deeper layers
- Perceptual loss (as loss function)
- Adam optimizer
- Use of loss as a metric for in-distribution (ID) or out-of-distribution (OOD) samples
Datasets:
[...]
Results published at:
# Autoencoders to detect manifestation shift in medical images
@inproceedings{freitas2025autoencoder,
author={SA Freitas, CA da Costa, GO Ramos},
booktitle={Simposio Brasileiro de Computação Aplicada a Saúde {SBCAS}, 2025, Porto Alegre, BRAZIL, June 9-13, 2025},
publisher = {{IEEE}},
title={Autoencoders to detect manifestation shift in medical images},
year={2025}
}