KMID : 1114620220190020060
|
|
Journal of the Korean Society for Breast Screening 2022 Volume.19 No. 2 p.60 ~ p.67
|
|
Real-Time Breast Ultrasound Artificial Intelligence Diagnosis using Dynamic Image Test Blocks
|
|
Jo Seo-Hyun
Kim Jae-Il Kim Chan-Ho Shin Ho-Kyung Kim Hye-Jung Yoon Jung-Hyun Jung Hae-Jung Kim Won-Hwa
|
|
Abstract
|
|
|
Purpose: In this study, we propose an AI-based decision support method that detects and differentiates lesions from given breast ultrasound images in real-time. We also evaluate the effectiveness of the proposed method using a keyblock-wise accuracy and computation time measurement.
Materials and Methods: A weakly supervised deep learning model for the real-time decision support was implemented and trained using more than 500,000 breast ultrasound images. To evaluate the real-time performance of the proposed model on dynamic ultrasound images, we collected six ultrasound videos from breast cancer patients at Kyungpook National University Chilgok Hospital. In the model evaluation, 25 key-frame blocks with 201 frames per block were
extracted using a landmark-based key-frame extraction algorithm. The block-wise labeling was performed by an experienced radiologist with reference to pathologic and radiologic reports. Diagnostic performance indicators were accuracy, AUC, sensitivity, and specificity. For the evaluation of real-time processing, the computation time per frame was measured in both hardware with GPU (HW Configuration 1) and hardware without GPU (HW Configuration 2).
Results: The key-frame blocks were labeled with 14 non-malignancy cases and 5 malignancy cases (6 cases were excluded), and each block contained 201 frame images. The diagnostic performance was 0.882 of accuracy, 0.917 of AUC and sensitivity/specificity was 1.0/0.833, respectively. The inference time was 0.03 seconds per frame in HW Configuration 1 and 0.06 seconds per frame in HW Configuration 2.
Conclusion: The developed model can support breast ultrasound diagnostics in real-time due to high diagnostic performances and fast inference time.
|
|
KEYWORD
|
|
Breast, Ultrasound, Cancer, Real-time, Dynamic, Artificial Intelligence (AI), Computer-aided Diagnosis (CAD)
|
|
FullTexts / Linksout information
|
|
|
|
Listed journal information
|
|
|
|