Improvement in detecting and localizing intracranial hemorrhage lesions using the active learning concept and probabilistic CAM heatmap

Abstract

Background Intracranial hemorrhage (ICH) treatment requires a prompt diagnosis based on a CT scan evaluation by a radiologist. Manually analyzing the images is complex and time-consuming. Deep-learning techniques have been successfully applied to assist radiologists with this task. We further improved the detection and localization of ICH lesions without subtype classification using the active learning concept and the Probabilistic CAM (PCAM) heatmap. Methods The train set of the second stage 2019-RSNA ICH data set was randomly separated into the train (712,803 slices), validate (20,000), and test (20,000) data sets and interpolated from 512 into 256, 1024, and 2048 image resolutions. All resolutions were scaled to RGB channels according to their subdural, brain, and bone windows and used in a DenseNet-121 2D-CNN classifier with the PCAM technique for localization using ICH labels. The point closest-to-(0,1) corner approach was used to find a cutoff point of the training data set on each image resolution and applied the cutoff value to the validate and test data sets to calculate performance metrics. Two neuroradiologists reevaluated the mismatched data between the RSNA labels and prediction outcomes on the test data set and reassigned labels when the evaluations agreed with the prediction outcomes. Both radiologists also evaluated the usefulness of the PCAM heatmap to localize ICH lesions into three categories: not useful (Poor), functional (Acceptable), and valuable (Good) grades. The improvement in ICH detection after reassigned RSNA labels was measured by AUROC and AUPRC. McNemar’s test determined whether image resolution would have a similar detection outcome with a significant level at P < 0.05. Results The study found that detecting ICH at 512 and 1024 resolutions gave a comparable performance and was significantly higher than at 256 and 2048 resolutions. However, their AUROC and AUPRC were still in the same range of 0.98 and 0.93, respectively. After reevaluation, 55% (471 from 859 FP) and 51% (114 from 225 FN) of the test data set were relabeled. Furthermore, using the same cutoff value, the AUROC and AUPRC can be increased by 1.1% (0.984 vs. 0.995) and 5.3% (0.932 vs. 0.981), respectively. The PCAM heatmaps obtained a Good grade of around 86%, 37% and 0% at 2048, 1024, and 512 resolutions, respectively. Conclusion Image resolution has a minor effect on altering the ICH detection performance, while reassessing the RSNA labels can significantly improve the performance. PCAM heatmaps can better localize the boundary regions of the ICH lesion at higher resolutions, with the best result in 2048.