Deep learning-based detection system for multiclass lesions on chest radiographs: comparison with observer readings
Sohee Park1, Sang Min Lee2, Kyung Hee Lee3
1Department of Radiology, University of Ulsan College of Medicine, Asan Medical Center, 88 Olympic-ro 43 Gil, Songpa-gu, Seoul, 138-736, South Korea.
European Radiology
|November 22, 2019
Summary
A deep learning-based detection (DLD) system demonstrated high feasibility and performance in identifying multiclass lesions on chest radiographs. This AI system outperformed human observers in both image-wise classification and lesion-wise detection tasks.
Area of Science:
- Radiology and Medical Imaging
- Artificial Intelligence in Healthcare
- Computer-Aided Diagnosis
Background:
- Chest radiography is a common imaging modality for diagnosing various thoracic abnormalities.
- Accurate detection and classification of multiple lesion types are crucial for effective patient management.
- Human observers' performance can vary based on experience and fatigue, necessitating complementary tools.
Purpose of the Study:
- To evaluate the feasibility and diagnostic performance of a deep learning-based detection (DLD) system for multiclass lesions on chest radiographs.
- To compare the DLD system's performance against that of human observers with varying experience levels.
- To assess the DLD system's capability in image-wise normal/abnormal classification and lesion-wise detection with pattern classification.
Main Methods:
- A dataset of 15,809 chest radiographs (7204 normal, 8605 abnormal) was utilized to develop the DLD system.
- The DLD system was trained on a large cohort, with a separate test set for evaluation.
- Diagnostic performance was assessed using area under the receiver operating characteristic curve (AUROC) for image-wise classification and figure of merit (FOM) for lesion-wise detection, compared against nine human observers.
Main Results:
- The DLD system achieved significantly higher performance than group-averaged human observations for both image-wise classification (AUROC, 0.985 vs. 0.958) and lesion-wise detection (FOM, 0.962 vs. 0.886).
- The DLD system outperformed all nine individual observers in lesion-wise detection.
- The system demonstrated superior performance across all lesion types, including nodules/masses and other abnormalities, with a lower false-positive fraction (0.11 vs. 0.19).
Conclusions:
- The deep learning-based detection system is feasible and highly effective for detecting and classifying multiclass lesions on chest radiographs.
- The DLD system exhibits excellent diagnostic performance, surpassing human observers and showing potential for integration into clinical workflows.
- The system's high accuracy in distinguishing normal from abnormal radiographs and its superior lesion detection capabilities highlight its value in improving radiological assessments.
