(주) 뷰노

(주) 뷰노

문의사항 남기기
KR EN
VUNO MED INC.
Publications
BACK TO LIST
RSNA 2025

A Deep Learning-based Microbleed Segmentation for ARIA-H Detection and Radiographic Severity Prediction in Out-of-Distribution Cases

  • Nov. 2025
  • by Saehyun Kim et. al.

Purpose

To evaluate a deep learning-based algorithm for automatic cerebral microbleed (CMB) segmentation in SWI MRI across severity levels. We also assess the algorithm’s performance on out-of-distribution cases with extreme CMB burden. Manual detection of CMBs is time-consuming, subjective, and suffers from inter-reader variability, affecting treatment decisions. This study addresses the need for objective CMB quantification tools with proven robustness across diverse patient populations.

 

Methods and Materials

We collected 565 SWI MRI scans (normal:136, mild:390, moderate:32, severe:7) from a single South Korean tertiary hospital. A neuroradiologist with 14 years of experience reviewed MRIs and labeled CMBs. The MRIs were split in a 3:1:1 ratio for training, validation, and testing. We additionally included 28 cases with >10 CMBs and 5 CADASIL cases as challenging out-of-distribution cases with extremely high microbleed burden (59-107 CMBs per case). An Attention U-Net with deep supervision was trained. Evaluation included both lesion-level and patient-level analysis across severity categories.

 

Results

The model achieved lesion-level metrics of 91.8% sensitivity, 54.7% precision, and 68.5% F1 score in detecting microbleeds on the test set, with false positives per scan of 1.06 (114 cases). For evaluation of out-of-distribution cases at the lesion level, the model maintained 81.2% sensitivity on cases with >10 CMBs (76.5% precision, 4.46 FP/scan in 28 cases), while CADASIL cases presented challenges with 69.8% sensitivity, though higher precision (86.3%, 9.8 FP/scan in 5 cases). At the patient level, the model achieved 77.2% sensitivity, 68.3% specificity, and 82.2% precision, with performance varying by severity (86.3%, 66.7%, and 97.0% sensitivity for mild, moderate, and severe cases respectively). The average processing time was 35.4s per scan using a single V100 GPU.

 

Conclusion

The study presents an effective algorithm for CMB segmentation on SWIs across all severity levels, achieving 91.8% sensitivity in test cases. It showed adaptability in challenging cases like CADASIL (69.8% sensitivity, 86.3% precision). Error pattern analysis revealed optimization opportunities, particularly for CMBs <1mm>

 

Clinical Relevance/Application

The deep learning approach provides efficient, objective quantification of CMBs as an assistive tool for clinicians in ARIA monitoring. The proposed method could reduce reading time, streamline clinical workflow, reduce inter/intra-rater inconsistency in MRI assessment among radiologists of varying experience levels, and guarantee more consistent treatment decisions.

Tag in

#medical_image

#VUNO Med®-DeepBrain®

이전글
Development and validation of a deep learning-based automatic segmentation and classification of cerebral white matter hyperintensities 2025. 12. 15
다음글
SemiSegECG: A Multi-Dataset Benchmark for Semi-Supervised Semantic Segmentation in ECG Delineation 2025. 11. 11
List

이 웹사이트는 원활한 서비스 제공을 위하여 당사 개인정보 처리방침에 명시된 쿠키를 사용합니다.

만일 당사의 쿠키 사용 정책에 대해 더 알고 싶으신 경우에는 개인정보 처리방침 을 참고하여 주시기 바랍니다.