(주) 뷰노

(주) 뷰노

문의사항 남기기
KR EN
VUNO MED INC.
Publications
BACK TO LIST
KATRDIC 2025 (한국결핵및호흡기학회)

Deep Learning-Based Automatic Detection of Mediastinal AbnormalLesions in Chest CT Using Anatomical Pseudo-Labeling

  • Oct. 2025
  • by Kyoung Min Moon et. al.

KATRDIC 2025 (한국결핵및호흡기학회) Oral presentation, Best presentation Award.

Background

Incidental detection of mediastinal abnormal lesions, including masses, cysts, and nodules, has increased with widespread chest computed tomography (CT) use. Accurate localization is essential for timely management but remains challenging due to anatomical complexity and variable lesion presentation.

 

Methods

We retrospectively collected 182 chest CT scans from Chung-Ang University Hospital: 91 normal and 91 with mediastinal abnormal lesions. Board-Certified 2 pulmonologists and 1 thoracic surgeon provided slice- and lesion-level annotations. We developed a deep learning framework combining slice-level classification (ResNet-18 backbone) and object detection (Faster R-CNN with ResNet-50-FPN), evaluated under four input settings: Total View, Lung View, Lesion View, and Lesion Object View. To address sparse annotations, we implemented an anatomical 3D pseudo-labeling strategy propagating bounding boxes across adjacent slices using morphological priors and Hounsfield Unit–based intensity tracking. Performance was compared with 8 board-certified physicians.

 

Results

ResNet-18 with expert-annotated slices achieved the highest classification performance (AUROC 0.9861, F1-score 0.9348). The proposed detection model with pseudo-labeling attained a mean average precision (mAP) of 0.712, exceeding the mean expert performance (mAP 0.698) and outperforming several individual physicians. Ablation studies confirmed that expert annotations and balanced random sampling of normal slices significantly enhanced performance, while patient-level data splitting revealed more realistic but lower metrics compared with case-level splitting.

 

Conclusions

Our deep learning model, augmented by anatomical 3D pseudo-labeling, achieved expert-level performance in detecting mediastinal abnormal lesions on chest CT and reduced annotation burden. The method demonstrated robustness across varied lesion morphologies and anatomical locations, supporting its potential as a clinical decision-support tool, particularly in settings with limited radiology expertise. Future work will validate the approach in larger, multi-institutional datasets for broader clinical adoption.

Tag in

#medical_image

이전글
Improved Deep Learning-Based Computer-Aided Diagnosis for Usual Interstitial Pneumonia Classification in Chest CT scan 2025. 10. 29
다음글
Enhancement of Prediction of In-Hospital Cardiac Arrest of Deep learning based Early Warning Score (DEWS) in a Single Tertiary Hospital 2025. 10. 25
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.