What question did this study set out to answer?

To develop an AI-based approach for predicting breast cancer-associated genetic mutations using digital mammography.

May 24, 2026Open Access

370P Genotype-phenotype correlations in a Japanese cohort with BRCA1/2 pathogenic variants

Key Points

To develop an AI-based approach for predicting breast cancer-associated genetic mutations using digital mammography.
Developed an AI-driven framework using convolutional neural networks (CNNs) on a clinical cohort of 100 breast cancer patients.
Evaluated multiple deep learning architectures and utilized ensemble strategies to enhance predictive performance.
Used class activation mapping for explainability and Monte Carlo dropout to quantify predictive uncertainty.
DenseNet-121 achieved 73.9% model accuracy, with ensembles improving this to 78.4% and weighted averaging to 82.6% accuracy with F1 score 0.83.
Genotype-phenotype correlations showed luminal tumors in non-mutated patients, while BRCA1 carriers exhibited younger high-grade triple-negative disease.

Abstract

Background: Breast cancer(BC) remains the most commonly diagnosed malignancy among women worldwide.Early identification of high-risk genetic mutations, particularly BRCA1/2, is critical for therapeutic decision-making and risk-reduction strategies.However, access to genetic testing is limited in resource-constrained settings due to cost and logistical barriers.We aim to develop a non-invasive, AIbased approach to predict breast cancer-associated genetic mutations using digital mammography (DM). Methods:We developed an AI-driven framework using convolutional neural networks (CNNs), pre-trained on the Mini-DDSM dataset to learn foundational imaging features and subsequently fine-tuned on a single-center clinical cohort of 100 breast cancer patients (50 with pathogenic mutations and 50 with negative genetic testing).Multiple deep learning architectures (DenseNet-121, ResNet-50, MobileNetV2, InceptionV3, and custom CNN) were evaluated.Ensemble strategies (stacking and weighted averaging) were employed to improve predictive performance.Explainability was assessed using class activation mapping (CAM)-based techniques, and predictive uncertainty quantified using Monte Carlo dropout.Results: DenseNet-121 achieved the highest individual model accuracy (73.9%).Ensembles further improved performance, with stacking achieving 78.4% accuracy and weighted averaging demonstrating the best generalizability (accuracy 82.6%, F1score 0.83).Explainability analyses consistently highlighted clinically relevant mammographic regions.Genotype-phenotype correlations were observed: nonmutated patients mostly exhibited luminal tumors, BRCA1 carriers were younger with high-grade triple-negative disease, and BRCA2/non-BRCA mutation carriers largely demonstrated intermediate-grade luminal type BC. Conclusions:This study demonstrates the feasibility of using AI-based analysis of DM to identify mutation-associated imaging patterns in BC.While not a substitute for genetic testing, this approach may serve as a cost-effective, non-invasive triaging tool to prioritize patients for genetic evaluation, in settings with limited access to molecular testing.

Bookmark

View Full Paper