# Predicting Breast Imaging-Reporting and Data System Classification of Palpable Breast Masses Using Ultrasound to Prioritize Mammography Queues

## TL;DR

This study developed a model using ultrasound features to prioritize mammography for palpable breast masses, improving efficiency and accuracy in breast cancer diagnosis.

## Contribution

A novel two-step ultrasound-based model for predicting BI-RADS classification to optimize mammography queue prioritization.

## Key findings

- The model achieved an AuROC of 0.9801 in step 1 and 0.9623 in step 2, showing excellent discrimination.
- The model had 88.5% accuracy with minimal overestimation and slight underestimation in BI-RADS 4–5 cases.
- The model's performance was validated internally with 200 bootstrap cycles, confirming minimal optimism.

## Abstract

Breast cancer is the leading cause of cancer death in women worldwide. Breast imaging, usually mammography and/or ultrasound, is classified using the Breast Imaging-Reporting and Data System (BI-RADS). At Lampang Hospital, mammography delays of up to 5 months postpone diagnosis in 40% of breast cancer cases. An urgent queue for palpable breast masses was introduced, but nearly half were benign, leading to inefficient prioritization. This study aimed to develop a two-step model based on high-risk ultrasound features and compare it with reference BI-RADS classifications.

This diagnostic prediction study collected retrospective data from Lampang Hospital between January 2021 and December 2023. Ultrasound images of 390 patients were independently reviewed by radiologists blinded to the reference BI-RADS classification. Stepwise multivariable risk difference regression analysis was applied to identify predictive characteristics from seven predefined ultrasound findings.

Three predictive characteristics were identified: shape, margin, and echo pattern. The two-step model showed excellent discrimination, with an area under the receiver operating characteristic curve (AuROC) of 0.9801 (95% CI, 0.9696–0.9907) in step 1 and 0.9623 (95% CI, 0.9411–0.9835) in step 2. Internal validation with 200 bootstrap cycles confirmed minimal optimism. Using prevalence-based cut points, the model achieved 88.5% accuracy, with 6.7% underestimation in BI-RADS 4–5 (predicted as 3) and overestimation not exceeding 3% in any category.

A two-step ultrasound-based model using shape, margin, and echo pattern demonstrated excellent discrimination as well as high accuracy, with slightly increased underestimation and minimal overestimation. This re-scheduling strategy optimizes mammography queue prioritization, but external validation is required before clinical implementation.