Abstract PO5-07-04: Machine learning models can help physicians reduce unnecessary ultrasound guided breast biopsies

Abstract Background Breast ultrasound is widely used as a diagnostic tool, the weakness of the method though is a high false positive rate, leading to unnecessary biopsies. In recent years, machine learning (ML) is gaining attention for its excellent performance in image-recognition tasks. The objective of this study was to train and validate ML models to predict malignancy of breast masses identified by ultrasound using clinical features of the patients, image information from the ultrasound reports and attributes extracted from the images. Methods We prospectively collected clinical and ultrasonographic attributes as well as the images from 927 lesions classified as BI-RADS 2, 3, 4a, 4b, 4c, 5 and 6 submitted to percutaneous biopsy in four institutions. Images in PNG format were loaded with OpenCV library and converted to gray scale. A total of nine attributes were extracted from the annotated area. We trained and validated five ML algorithms (histogram gradient boosting, logistic regression, a pipeline of these two classifiers, stacking ensemble of several classifiers and lightgbm), randomly sampling 80% of the dataset for training and 20% for validation using a cross fold strategy. We trained and tested the models with five different combinations of the data collected and compared the performances. The model with the lowest mean errors was selected and the threshold tuned to minimize the cases that the model classified as benign and were truly cancer (“false not cancer”). Image preprocessing, attribute extraction, and traditional machine learning was performed in Python. Results The clinical characteristics of the patients and ultrasonographic attributes of the lesions withdrawn from the exam reports are described in Table 1. The biopsy results distribution across the BI-RADS classification of the ultrasound report is shown in Table 2 and are in accordance with the literature. Among the five tested models, the logistic regression was the one with the lowest mean errors and was selected to tune the thresholds minimizing “false not cancer” cases. Our target was up to 2% of this error to stay within the BI-RADS 3 classification and thus dismiss biopsy. Table 3 describes five different combinations of data used to construct the models. All these combinations had 0.5% of "false not cancer”, way below the 2% threshold from the BI-RADS 3 category. The mean error of the models classifying benign lesions as cancer and inducing unnecessary biopsies varied from 22.7% to 23.3% depending on the combination of data used. Conclusions Machine learning models can potentially help physicians avoid unnecessary ultrasound guided breast biopsies using unsophisticated clinical features and attributes from the images missing 0.5% of cancer cases. Table 1. Clinical characteristics of the included patients and ultrasound features of the breast lesions. Table 2. Histology (malignant or benign) distribution across BI-RADSⓇ classification. *We had five lesions missing this data. Table 3. Mean performances of the five models and parameters used to build them. Citation Format: Isabela Buzatto, Danilo Carlotti, Nilton Onari, Ana Luiza Faim, Sarah Recife, Licerio Miguel, Ruth Bonini, Liliane Silvestre, Lucas Figueira, Daniel Guimarães Tiezzi. Machine learning models can help physicians reduce unnecessary ultrasound guided breast biopsies abstract. In: Proceedings of the 2023 San Antonio Breast Cancer Symposium; 2023 Dec 5-9; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2024;84(9 Suppl):Abstract nr PO5-07-04.