Tumor Detection in Brain MRIs by Computing Dissimilarities in the Latent Space of a Variational AutoEncoder

Abstract

The ability to automatically detect anomalies in brain MRI scans is of great importance in computer-aided diagnosis. Unsupervised anomaly detection methods work primarily by learning the distribution of healthy images and identifying abnormal tissues as outliers. We propose a slice-wise detection method which first trains a pair of autoencoders on two different datasets, one with healthy individuals and the other one with images of normal and tumoral tissues. Next, it classifies slices based on the distance in the latent space between the enconding of the image and the encoding of the reconstructed image, obtained through the autoencoder trained on healthy images only. We validate our approach with a series of preliminary experiments on the HCP and BRATS-15 datasets.

Improved Slice-wise Tumour Detection in Brain MRIs by Computing Dissimilarities between Latent Representations

“…In this paper, we have presented novel and improved results on anomaly detection for brain MRIs based on two Variational Au-toEncoders (VAEs). Compared to our previous results [1], several changes have been made that have conducted to an improvement of the performance. A newer version of the BRATS dataset (BRATS-2018) has been used in training, together with a larger subset of the HCP dataset containing 214 individuals.…”

“…As in [1], we have chosen in Algorithm 1 the distance function d to be the norm of the difference between the means obtained with VAE for the encodings of x and x h . The threshold d * used to classify images from the test set of BRATS has been computed as a percentile of the distances computed on the validation set of HCP.…”

“…Table 1: Area Under the ROC Curve, accuracy and F1 score for BRATS-2018 (test set, averaged over 5 folds) for L 2 distances computed in the input space for different VAEs versus L 2 of the means in the latent space (our method [1]). Th.…”

“…In our previous work [1], we have designed a slice-wise tumour detection algorithm based on Variational AutoEncoders (VAEs) for tumour detection in brain MRIs from two publicly available datasets, HCP [16] and BRATS-2015 [9,11]. Our algorithm is based on the computation of a distance, or more generically a dissimilarity function, in the space of the approximate posteriors of a VAE, trained on both healthy (or normal) and tumoural tissues.…”

“…Classification of brain MRIs from[1] Input: Let d be a distance defined over the latent space of a VAEData: MRI-H: dataset of MRI slices of healthy individuals Data: MRI: dataset of MRI slices of individuals which may have tumoural tissues 1 Let VAE-H be a VAE trained on MRI-H 2 Let VAE be a VAE trained on MRI 3 Let x h = VAE-H.rec(x ) be the reconstruction of x through VAE-H 4 Let VAE.enc(x ) be the encoding of x through VAE 5 Let d x = d (VAE.enc(x ),VAE.enc(x h )) 6 Compute the distribution of the distances d x , with x in the validation set of MRI-H and let d * be a threshold selected based on a percentile 7 Compute d x , with x in the test set of MRI, and classify the slice as healthy if d x < d *…”