Intracranial hemorrhage detection: how much can windowing, transfer learning and data augmentation affect a deep learning model’s performance?

Abstract

The application of Deep Neural Networks to detect intracranial hemorrhage from computed tomography images has been widely used in clinical medicine. In general, building these solutions combines three types of approaches: Preprocessing, Transfer Learning, and Data Augmentation. This study’s goal was to measure the contribution of each of these approaches and thus highlight which approach has more margin of improvement. An experimental study was conducted on a public dataset containing computerized tomography images. The comparison used the stratified ten-fold cross-validation process to set confidence intervals evaluating performance measured by the area under the receiver operating characteristic curve. The ResNet-50 was the deep learning model selected. The results showed that all the approaches raise the generalization power when applied in isolation, and Data Augmentation offers the most significant gain to the baseline. The experiment also showed an opportunity to improve the detection of intracranial hemorrhage by applying new preprocessing techniques since this was the approach that showed the smallest increase in discriminatory power among the investigated approaches. The paired Wilcoxon’s Signed-Rank Test showed that not all the differences were statistically significant, with a confidence level of 95%.