TexVar: Um Autoencoder Variacional para representação e interpretação de texturas mamográficas

• Lucca Lemos Lago PUC Minas
• Alexei Manso Correa Machado PUC Minas / UFMG

DOI: https://doi.org/10.5753/sbcas.2023.229644

Abstract

The classification of breast density through X-ray exams is still considered the main screening mechanism for the early detection of breast cancer, since the fibroglandular tissue can hide initial tumors. The objective of this article is to propose an interpretable model for the classification of breast densities in mammographic images. The proposed architecture consists of a variational autoencoder (VAE) composed by convolutional and dense layers with a latent space of 32 variables. The model allows for explaining the meaning of the latent variables through the texture transition between the 4 classes of the Breast Image Reporting and Data System (BI-RADS) scale. Experiments using the IRMA public database composed of 5024 images demonstrated the ability of the VAE to reduce the dimensionality of the problem to a space where the most discriminating variables can be visually interpreted, supporting computer-aided diagnosis.

References

Habaragamuwa, H., Oishi, Y., e Tanaka, K. (2021). Achieving explainability for plant disease classification with disentangled variational autoencoders. CoRR, abs/2102.03082. Retrieved from https://arxiv.org/abs/2102.03082

INCA. (2021, Set). Estatísticas de câncer. https://www.inca.gov.br/numeros-de-cancer. (acessado em: 15/10/2021)

Kingma, D. P., e Welling, M. (2013). Auto-encoding variational bayes. arXiv. Retrieved from https://arxiv.org/abs/1312.6114 doi: 10.48550/ARXIV.1312.6114

Liu, W., Li, R., Zheng, M., Karanam, S., Wu, Z., Bhanu, B., . . . Camps, O. (2020). Towards Visually Explaining Variational Autoencoders. Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 8639–8648.

Oliveira, J. E., Gueld, M. O., Araújo, A. d. A., Ott, B., e Deserno, T. M. (2008). Toward a standard reference database for computer-aided mammography. In Medical imaging 2008: Computer-aided diagnosis (Vol. 6915, p. 69151Y).

Picard, R., Graczyk, C., Mann, S., Wachman, J., Picard, L., Negroponte, N., e Campbell, L. (1995, Mar). Retrieved from https://vismod.media.mit.edu/pub/VisTex/

Schockaert, C., Macher, V., e Schmitz, A. (2020). VAE-LIME: Deep Generative Model Based Approach for Local Data-Driven Model Interpretability Applied to the Ironmaking Industry.

Utkin, L., Drobintsev, P., Kovalev, M., e Konstantinov, A. (2021). Combining an auto encoder and a variational autoencoder for explaining the machine learning model predictions. Conference of Open Innovation Association, FRUCT , 2021-Janua. doi: 10.23919/FRUCT50888.2021.9347612

Uzunova, H., Ehrhardt, J., Kepp, T., e Handels, H. (2019). Interpretable explanations of black box classifiers applied on medical images by meaningful perturbations using variational autoencoders. In E. D. Angelini e B. A. Landman (Eds.), Medical imaging 2019: Image processing (Vol. 10949, pp. 264 – 271). SPIE. Retrieved from https://doi.org/10.1117/12.2511964 doi: 10.1117/12.2511964

Wei, R., e Mahmood, A. (2021). Recent advances in variational autoencoders with representation learning for biomedical informatics: A survey. IEEE Access, 9, 49394956. doi: 10.1109/ACCESS.2020.3048309