Detection of Tumours in Brain MRIs with Variational AutoEncoders

Alexandra-Ioana Albu, Alina Enescu, Luigi Malagò

Full text: https://drive.google.com/file/d/1H1XbrqCwdnM0_xH_SW1d73Mhoo02f1MY/view

Find on Google Scholar

---

Abstract

In presence of negative and unlabelled data, anomaly detection for brain MRIs can be framed as a semi-supervised task, by comparing the latent representations of an image and of its reconstruction through a model trained only on healthy individuals. Methods which employ the reconstruction of images through an autoencoder strongly depend on the image resolution and the capability of the model to reproduce relevant details in the reconstructions. In this extended abstract we show how the performance of slice-wise tumour detection can be improved by training more powerful models on higher resolution images, as well as by the use of a perceptual loss as a regularizer during training.

---

---