Breast Cancer, Breast Neoplasms
Conditions
Brief summary
This is a retrospective, fully-crossed, multi-reader, multi-case (MRMC) study to evaluate the effectiveness of 'CadAI-B Dx' (CadAI-B) for decision support in breast ultrasound. The study compares the diagnostic performance of readers interpreting breast ultrasound images with and without the aid of CadAI-B. A total of 797 patient cases will be included, comprising 350 cases with a confirmed diagnosis of malignancy and 447 cases with a confirmed benign diagnosis. Sixteen readers will participate in the study to evaluate the device.
Detailed description
The study utilizes a crossover design where all readers independently review all cases. The control arm consists of a reading session where participating readers independently review cases without the assistance of the CadAI-B device (unaided reading). The experimental arm involves reading with CadAI-B assistance (AI-aided reading). To minimize potential bias, a washout period of four weeks will be maintained between the unassisted and assisted reading sessions for each reader. The primary hypothesis is that CadAI-B assistance significantly improves overall reader performance in breast ultrasound interpretation, as measured by the area under the Localization Receiver Operating Characteristic (LROC) curve (AULROC).
Interventions
DEVICECadAI-B Dx
CadAI-B Dx is a Software as a Medical Device (SaMD) designed to assist physicians by providing Computer-Aided Detection (CADe) and Diagnosis (CADx) capabilities in breast ultrasound interpretation. The software automatically processes the image to identify suspicious regions (Lesion Detection) and provides a quantitative malignancy score (CadAI-Score) mapped to a corresponding BI-RADS Category. It also analyzes lesion size and BI-RADS lexicon descriptors.
Sponsors
BeamWorks Inc.
Study design
Allocation
RANDOMIZED
Intervention model
CROSSOVER
Primary purpose
DIAGNOSTIC
Masking
QUADRUPLE (Subject, Caregiver, Investigator, Outcomes Assessor)
Intervention model description
Retrospective, fully-crossed, Multi-Reader Multi-Case (MRMC) study.
Eligibility
Sex/Gender
FEMALE
Age
22 Years to No maximum
Healthy volunteers
No
Inclusion criteria
1. Women aged 22 years or older at the time of the breast ultrasound examination. 2. Standard B-mode breast ultrasound images available for analysis. 3. Lesions must have a confirmed diagnosis based on one of the following reference standards: 3-1. Malignant Cases: Diagnosis of breast cancer confirmed through biopsy or surgery. 3-2. Benign Cases: Confirmed as benign through biopsy or surgical excision, or confirmed as having no evidence of malignancy for at least 2 years of follow-up.
Exclusion criteria
1. mages containing modes other than standard B-mode ultrasound, such as Doppler, elastography, or other annotations/overlays. 2. Patients who have breast implant(s). 3. Patients suffering from significant breast trauma or mastitis at the time of the breast ultrasound examination. 4. Images of post-surgical resection sites. 5. Images where multiple lesions are present within a single 2D ultrasound image
Design outcomes
Primary
| Measure | Time frame | Description |
|---|---|---|
| Area Under the Localization Receiver Operating Characteristic (LROC) Curve (AULROC) | Through study completion, approximately 2 months | The difference in reader performance between the unaided and AI-aided sessions in breast ultrasound interpretation. LROC reflects both detection accuracy and localization precision. The primary hypothesis is that the mean AULROC of all readers for the AI-aided reading mode is greater than that for the unaided reading mode. |
Secondary
| Measure | Time frame | Description |
|---|---|---|
| Positive Predictive Value (PPV) | Through study completion, approximately 2 months | Comparison of PPV between unaided and AI-aided sessions. The PPV will be adjusted for disease prevalence in the target population to reflect real-world clinical practice. |
| Negative Predictive Value (NPV) | Through study completion, approximately 2 months | Comparison of NPV between unaided and AI-aided sessions. The NPV will be adjusted for disease prevalence in the target population. |
| Sensitivity | Through study completion, approximately 2 months | Comparison of the average sensitivity of the readers between the unaided and AI-aided sessions. |
| Reading Time | Through study completion, approximately 2 months | Comparison of the average reading time per case between the unaided and AI-aided sessions to assess if the AI system improves the efficiency of interpretation. |
| AI-Ground Truth Agreement | Through study completion, approximately 2 months | Assessment of the agreement between the AI system's outputs and the ground truth. This includes agreement on BI-RADS categories and descriptors (using Kappa statistics) and lesion size measurements (using Intraclass Correlation Coefficient). |
| Inter-reader Agreement | Through study completion, approximately 2 months | Evaluation of the consistency of interpretations among different readers. Inter-reader agreement for BI-RADS category and descriptors assignments will be compared between sessions using Kappa statistics. |
Countries
South Korea
Outcome results
None listed