Prior knowledge-based precise diagnosis of blend sign from head computed tomography

Intracranial hemorrhage detection remains a critical challenge in emergency care. While head CT scans are widely used, interpreting subtle signs like the blend sign requires specialized expertise. Prior research has shown that automated systems can support radiologists, but limitations persist in distinguishing complex features. No prior work had resolved the challenge of integrating hemorrhage location as prior knowledge into detection models. This gap motivated the development of a method that uses auxiliary tasks to guide attention toward hemorrhage regions. Existing models often struggle with inaccurate annotations, which can reduce diagnostic accuracy. That uncertainty drove the need for a self-knowledge distillation approach to improve robustness. The blend sign, in particular, is difficult to classify due to overlapping features with normal ICH. This paper addresses these limitations by combining classification and object detection in a novel framework.

This study aimed to improve the accuracy of blend sign detection in head CT scans. The blend sign is a subtle indicator of intracranial hemorrhage that can be easily misclassified. The researchers proposed a method that incorporates hemorrhage location as prior knowledge. This approach uses an auxiliary object detection task to guide the model's focus. The study also sought to address the issue of inaccurate annotations in training data. A self-knowledge distillation strategy was introduced to refine model predictions. The goal was to create a system that could assist less-experienced interpreters. The researchers tested their method on a dataset of 1749 non-contrast head CT scans.

The researchers designed a dual-task framework combining classification and object detection. The classification task identified the presence of intracranial hemorrhage. The object detection task localized hemorrhage regions within the CT scans. This localization served as prior knowledge to guide the model's attention. The auxiliary task was trained alongside the main classification task. A self-knowledge distillation strategy was implemented to refine model outputs. This strategy helped reduce the impact of annotation errors in the training data. The model was tested on a dataset of 1749 non-contrast head CT scans. The dataset included three categories: no ICH, normal ICH, and blend sign.

The proposed method outperformed existing approaches in blend sign detection. The model achieved higher accuracy than baseline methods in distinguishing the blend sign. The use of object detection as an auxiliary task improved attention to hemorrhage regions. The self-knowledge distillation strategy reduced the effects of annotation errors. The model demonstrated strong performance across all three categories. The researchers reported improved sensitivity and specificity compared to prior models. The method showed potential for use in clinical settings with less-experienced interpreters. The results suggest that the model could reduce radiologists' workload and improve diagnostic efficiency.

The study demonstrated that incorporating hemorrhage location as prior knowledge improves blend sign detection. The dual-task framework with object detection and classification enhanced model performance. The self-knowledge distillation strategy helped address annotation inaccuracies. The method showed promise for supporting less-experienced head CT interpreters. The researchers propose that this approach could reduce radiologists' workload. The results suggest the model could improve diagnostic efficiency in clinical settings. The method may be useful for distinguishing the blend sign from normal ICH. The study may suggest that integrating prior knowledge into detection models is beneficial.