Overview

Schema Version

https://atlas.rsna.org/schemas/2025-11/model.json

Name

AI-assisted Analysis to Facilitate Detection of Humeral Lesions on Chest Radiographs

Link

https://github.com/kskim-phd/FPAR

Indexing

Keywords: Artificial Intelligence, Conventional Radiography, Humerus, Machine Learning, Shoulder, Tumor

Content: CH, MK

Author(s)

Harim Kim

Kyungsu Kim

Seong Je Oh

Sungjoo Lee

Jung Han Woo

Jong Hee Kim

Yoon Ki Cha

Kyunga Kim

Myung Jin Chung

Organization(s)

Department of Radiology, Samsung Medical Center, Sungkyunkwan University School of Medicine, Seoul, South Korea

Medical AI Research Center, Samsung Medical Center, Seoul, South Korea

Department of Data Convergence and Future Medicine, Sungkyunkwan University School of Medicine, Seoul, South Korea

Samsung Advanced Institute for Health Sciences & Technology (SAIHST), Sungkyunkwan University, Seoul, South Korea

Version

1.0

Contact

ude.ukks@gnuhcjm

Funding

Korea Medical Device Development Fund (no. 202011B08-02, KMDF_PR_20200901_0014-2021-02), Technology Innovation Program (no. 20014111) funded by MOTIE (Korea), and Future Medicine 20*30 Project of Samsung Medical Center (no. SMX1210791).

Ethical review

Institutional Review Board approval obtained; informed consent waived (Samsung Medical Center IRB no. 2022-08-099).

Date

Updated: 2024-02-15

Published: 2024-03-06

Created: 2023-03-25

References

[1] Kim H, Kim K, Oh SJ, Lee S, Woo JH, Kim JH, Cha YK, Kim K, Chung MJ. "AI-assisted Analysis to Facilitate Detection of Humeral Lesions on Chest Radiographs". Radiology: Artificial Intelligence. 2024 May;6(3):e230094.. 2024-03-06. doi:10.1148/ryai.230094. PMID: 38446041. PMCID: PMC11140509.

Model

Architecture

Convolutional Neural Network with automated humerus segmentation preprocessing and a training regularizer termed false-positive activation area reduction (FPAR); Grad-CAM used for visualization. Fivefold cross-validation ensemble (averaged probability vectors).

Availability

Code available at https://github.com/kskim-phd/FPAR

Clinical benefit

Assists radiologists in detecting humeral tumors on chest radiographs, improving sensitivity, specificity, and accuracy for humeral lesion detection.

Clinical workflow phase

Clinical decision support during image interpretation/readout.

Decision threshold

0.35 (selected to maximize sensitivity among thresholds with false-positive rate < 0.05; used for reader comparison).

Degree of automation

Fully automated preprocessing (humerus segmentation and region extraction) and automated binary classification per humerus; outputs provided to readers as assistive results.

Indications for use

Detection of primary or metastatic humeral tumors on chest radiographs in patients undergoing chest radiography, intended for use as a supportive diagnostic tool in clinical environments such as cancer centers and tertiary hospitals.

Input

Chest radiographs (PA or AP, fixed or portable) with both humeral heads visible; automatically segmented humerus regions used as input to the diagnostic model.

Instructions

Use as an assistive tool: review AI binary outputs per humerus and associated visualization to potentially adjust interpretations. Readers in the study reviewed CRs, then reviewed AI outputs and could change their answers.

Limitations

Single-institution, single-ethnicity study; no external testing; relatively low sensitivity for some small/low-visibility tumors on CR; threshold preselected (0.35) for comparisons; limited number of primary humeral tumors in training; multiple images per patient may introduce bias (sensitivity/specificity robust in sensitivity analysis); general population prevalence is low, so false positives may be impactful; radiologist test setting may not reflect real-world practice.

Output

CDEs: RDE2804, RDE2090

Description: Binary classification (tumor present/absent) for each humerus on a chest radiograph, with heatmap visualization for localization.

Recommendation

Use as a supportive diagnostic tool to alert radiologists to potential humeral abnormalities, particularly in cancer-rich populations (cancer centers, tertiary hospitals); not recommended as a stand-alone diagnostic algorithm.

Regulatory information

Authorization status: Not reported

Reproducibility

Fivefold cross-validation with holdout testing; large normal-only holdout set for specificity; code repository available for the FPAR method.

Use

Intended: Detection and diagnosis

Out-of-scope: Detection, Diagnosis

User

Intended: Radiologist, Other

Out-of-scope: Patient, Other