Random forest with preoperative core biopsy categories: a novel method for refining ultrasonic Breast Imaging...

Area of Science:

• Radiology
• Machine Learning
• Oncology

Background:

• Many benign breast lesions are classified as BI-RADS category 4, leading to unnecessary biopsies.
• Improving diagnostic accuracy for breast lesions is crucial to avoid invasive procedures.

Purpose of the Study:

• To develop a core needle biopsy category (CBC) prediction model to enhance BI-RADS classification.
• To reduce the number of unneeded biopsies for benign breast lesions.

Main Methods:

• A retrospective study analyzed clinical and ultrasonic features of 1,082 female patients with solid breast tumors.
• Five machine learning algorithms were used to develop CBC prediction models, with random forest (RF) selected as optimal (AUC=0.943).
• The optimal model adjusted BI-RADS categories for lesions, downgrading category 3 and 4A lesions based on predicted CBC.

Main Results:

• The RF model accurately adjusted BI-RADS categories for 1,185 lesions.
• For BI-RADS category 3 lesions, 90.5% were correctly downgraded.
• For BI-RADS category 4A lesions, 89.2% were correctly downgraded, and 10.8% were correctly upgraded.

Conclusions:

• A machine learning-based CBC predictive model can effectively adjust BI-RADS categories 3 and 4A.
• This model has the potential to significantly decrease unnecessary breast biopsies.