# Diagnostic performance and confidence of an optimized deep-learning algorithm for the detection of intracranial hemorrhages

## TL;DR

An optimized AI algorithm detects brain bleeds in CT scans with high accuracy, comparable to radiologists, and may help reduce missed cases.

## Contribution

The paper introduces an optimized deep-learning algorithm that improves ICH detection accuracy and confidence compared to its initial version and radiologists.

## Key findings

- The current AI prototype detected ICH with 93.9% sensitivity and 96.1% specificity.
- Combining AI with radiologists may reduce missed ICH cases and improve diagnostic quality.
- A confidence score threshold of 60% improved AI performance significantly.

## Abstract

To evaluate the performance of an optimized deep-learning-based algorithm (AI) for the detection and subtyping of intracranial hemorrhage (ICH) in non-contrast cranial CT (cCT).

CCTs performed between 2020 and 2022 were processed using a pre-trained 3D-neural-AI to detect ICH. The current version was compared to the initial version (ICH) and the radiological report (ICH, subtypes) regarding diagnostic accuracy. A consensus of the radiological report and an additional reading by a radiologist (7 years of experience) served as the ground truth. We investigated the AI-generated confidence score as a threshold for clinical usage.

In the total cohort of 2960 cCTs (ICH prevalence 10.5%), the current AI prototype detected ICH with high sensitivity (93.9%) and specificity (96.1%). This resulted in an accuracy of 95.9% and a negative predictive value (NPV) of 99.3%, including 12 ICH-positive cases that were initially missed by the interpreting radiologists. Subtyping results were comparable between the AI and the radiologists. In the cohort processed with both prototypes (n = 996), the results of the current AI were slightly lower (sensitivity 88.3%; accuracy 94.4%; NPV 98.6%), yet it still outperformed those of the initial version (sensitivity 77.7%, accuracy 95.5%, NPV 97.4%), resulting in eight fewer false negatives and eleven additional true positives. A confidence score of 60% was considered a useful threshold, resulting in a significant increase of AUC (p = 0.018).

The current AI algorithm achieves high diagnostic accuracy and negative predictive value. Combining AI-driven analysis with radiologists’ expertise may improve the overall performance and reduce the number of missed ICHs.

The dual use of AI as a control and triage tool can reduce radiology workload. Our results show AI reliably supports standardized exams with diagnostic quality comparable to radiologists, while transparent output enhances clinical acceptance and integration.

Potentially life-threatening intracranial hemorrhages are time-critical and need accurate detection on non-contrast cranial CT.The optimized algorithm achieved high diagnostic accuracy for ICH detection similar to radiologists.The combination of radiologists and AI may improve the efficiency and diagnostic quality of ICH detection.

Potentially life-threatening intracranial hemorrhages are time-critical and need accurate detection on non-contrast cranial CT.

The optimized algorithm achieved high diagnostic accuracy for ICH detection similar to radiologists.

The combination of radiologists and AI may improve the efficiency and diagnostic quality of ICH detection.