Development and validation of a self-attention network-based algorithm to detect mediastinal lesions on computed...

Area of Science:

• Artificial Intelligence in Medical Imaging
• Deep Learning for Diagnostic Imaging
• Radiology and Medical Diagnostics

Background:

• Mediastinal lesion diagnosis using computed tomography (CT) presents challenges due to diverse presentations.
• Accurate identification of these lesions is crucial for timely and effective patient management.
• Existing diagnostic methods can be time-consuming and require significant radiologist expertise.

Purpose of the Study:

• To develop and evaluate a novel self-attention network-based algorithm for automated mediastinal lesion detection on CT images.
• To assess the algorithm's performance in terms of accuracy, sensitivity, and confidence in lesion identification.
• To compare the proposed algorithm's efficacy against established deep learning models like faster region-based convolutional neural network (R-CNN).

Main Methods:

• Utilized two large-scale open datasets: NIH DeepLesion and MICCAI 2022 MELA Challenge.
• Trained a self-attention network using 921 abnormal CT images for pretraining and 880 for model training/validation.
• Evaluated performance using average precision (AP) and confidence scores, comparing sensitivity against faster R-CNN.

Main Results:

• The self-attention network achieved an 89.3% average precision (AP) for mediastinal lesion detection.
• The model demonstrated high confidence (>0.8) in identifying large lesions.
• Achieved a nearly 2% performance improvement in Competition Performance Metric (CPM) over faster R-CNN, with high sensitivity and low false-positive rates.

Conclusions:

• The developed self-attention network shows excellent performance in detecting mediastinal lesions on CT scans.
• This AI tool has the potential to significantly reduce radiologist workload and expedite reporting times.
• Implementation can lead to improved diagnostic accuracy and efficiency in clinical practice.