Evaluation of TaiHao Breast Ultrasound Diagnosis Software

Not Applicable

Completed

Interventions: Diagnostic Test: Reader Group Y - Session 2
Diagnostic Test: Reader Group Y - Session 1
Diagnostic Test: Reader Group X - Session 2
Diagnostic Test: Reader Group X - Session 1

Brief Summary: The BR-USCAD DS Module is a computer-assisted detection and diagnosis software based on a deep learning algorithm. This retrospective, fully-crossed, multi-reader, multi-case (MRMC) study aims to compare the performances of readers without and with the aid of the Breast Ultrasound Image Reviewed with Assistance of Computer-Assisted Detection and Diagnosis System (BR-USCAD DS) in interpreting breast ultrasound images of lesions.

Recruitment & Eligibility

Inclusion Criteria

B-mode breast ultrasound image
Female, age 21 or older
Breast lesion images acquired before a biopsy or surgery - these images were retrospectively collected with histology report.
Non-biopsied benign lesions with negative follow-up for a minimum of 24 months
At least two orthogonal views of a lesion

Exclusion Criteria

Breast lesion images acquired after biopsy or surgery.
Any breast surgeries or interventional procedures in the 12 months prior to ultrasound imaging
Case demonstrating administrative or technical errors
Multiple lesions in one 2-D ultrasound image
Breast ultrasound images with Doppler, elastography, or other overlays present
Case with less than 2-year follow-up and without biopsy confirmation

Arm && Interventions

Group	Intervention	Description
Second session: review with CADx first and then manual review	Reader Group Y - Session 2	At least 4 weeks after first session for memory washing out. Rader Group Y interpret the "Dataset A" cases in different random order without any assistance of AI first, and then interpret the "Dataset B" cases in different random order with TaiHao AI system.
First session: review with CADx first and then manual review	Reader Group Y - Session 1	Reader Group Y interpret the "Dataset A" cases in different random order with TaiHao AI system first, and then interpret the "Dataset B" cases in different random order without any assistance of AI.
Second session: manual review first and then review with CADx	Reader Group X - Session 2	At least 4 weeks after first session for memory washing out. Reader Group X interpret the "Dataset A" cases in different random order with TaiHao AI system first, and then interpret the "Dataset B" cases in different random order without any assistance of AI.
First session: manual review first and then review with CADx	Reader Group X - Session 1	Reader Group X interpret the "Dataset A" cases in different random order without any assistance of AI first, and then interpret the "Dataset B" cases in different random order with TaiHao AI system.

Primary Outcome Measures

Name	Time	Method
Comparing the Area Under the LROC Curve	10 weeks	The area under the LROC curve (AUC_LROC) on the diagnosis of benign and malignant lesions was computed and compared for Aim 1 (baseline) and Arm 2 (with BU-CAD assistance) studies.

Secondary Outcome Measures

Name	Time	Method
The Sensitivity, Specificity, PPV, and NPV Were Computed and Compared for Aim 1 (Baseline) and Arm 2 (With BU-CAD Assistance) Studies.	10 weeks	The mean sensitivity, specificity, PPV, and NPV of 16 readers were calculated and compared between the aided and unaided sessions using McNemar's test.
The Reading Time Was Computed and Compared for Aim 1 (Baseline) and Arm 2 (With BU-CAD Assistance) Studies.	10 weeks	Each reader's reading time of a case was automatically recorded by the BU-CAD system. the average reading times of 16 readers with and without outlier reading times were compared between the aided and unaided sessions using the paired t-test.