A Comparative Study of Deep Learning-Based Anomaly Detection for Intracranial Hemorrhage in Brain CT

Joon Hyuk Lee, Hyeon Jun Park, Tae-Seong Kim, Won Hee Lee

Keywords: Anomaly Detection, Deep Learning, Intracranial Hemorrhage, Brain, CT

TL;DR: A systematic benchmark of 11 deep learning anomaly detection methods for intracranial hemorrhage detection in brain CT favors reconstruction-based over self-supervised approaches.

Registration Requirement: Yes

Abstract: Deep learning-based anomaly detection (AD) offers a label-efficient approach to intracranial hemorrhage (ICH) detection in brain CT, yet systematic comparisons of AD methods on this clinically critical task remain scarce. We benchmark 11 reconstruction-based and self-supervised deep learning AD methods on the CQ500 dataset under consistent experimental conditions. Our results reveal a substantial performance gap between the two paradigms: reconstruction-based methods achieve up to AUC 90.8, while self-supervised approaches peak at AUC 64.7, suggesting that synthetic anomaly strategies designed for natural images do not transfer well to brain CT. Among reconstruction-based methods, uncertainty-aware modeling consistently provides the most discriminative anomaly scores. This study presents a systematic benchmark of deep learning AD methods for ICH detection, offering concrete guidance for method selection in automated brain CT screening.

Visa & Travel: No

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 80