Abnormality Detection of Humerus Radiographs Using an Uncertainty Prediction Autoencoder Model

Abstract

This study aimed to investigate the performance of an uncertainty prediction autoencoder (UPAE) compared to a vanilla autoencoder (VAE) on the humerus. VAE is a standard neural network that compresses data into a lower-dimensional representation and recreates the image from the lower-dimensional representation, while UPAE is a type of autoencoder that is designed to estimate the uncertainty of its predictions. The models were trained on a dataset of humerus radiographs and evaluated on their ability to reconstruct the original data and compute for the reconstruction error. Comparing the results of the UPAE and VAE models shows that the addition of the Uncertainty Prediction does improve the performance of the models in terms of prediction anomalies. The best UPAE model scored 0.743 AUC and 0.691 F1, while the leading VAE model had 0.741 AUC and 0.654 F1. However, when tested on a local dataset, UPAE scored 0.543 AUC and 0.631 F1, while VAE outperformed with 0.691 AUC and 0.667 F1.

Ähnliche Arbeiten