Clinically Robust Deep Learning for Contrast-Enhanced Mammography: Multicenter Evaluation Across Convolutional Neural Network Architectures | JoVE Visualize

Area of Science:

Radiology and Medical Imaging
Artificial Intelligence in Healthcare
Biomedical Engineering

Background:

Deep learning models for breast lesion classification in contrast-enhanced mammography (CEM) face challenges with robustness and reliability across different datasets.
Anatomically constrained preprocessing and deep learning architecture selection are key factors influencing model performance.

Purpose of the Study:

To investigate the impact of anatomically constrained preprocessing (breast-mask segmentation) versus original DICOM images on deep learning model performance for CEM breast lesion classification.
To evaluate various deep learning architectures, including CNNs, attention-based networks, and Transformers, for their effectiveness in this task.

Main Methods:

A retrospective multicenter study combined 300 patient CEM images with 1003 public images (total 1120 cases).
Automatic breast segmentation was performed using the LIBRA framework to create breast-mask images.
Eleven deep learning models were trained and evaluated on both original DICOM and breast-mask inputs, assessing performance metrics like AUROC and balanced accuracy.

Main Results:

Models trained with breast-mask images consistently outperformed those trained on original DICOM images, with AUROC improvements of +0.06 to +0.21.
ResNet50 with breast-mask input achieved the highest performance (AUROC=0.931), further improved after optimization (balanced accuracy=0.886).
Classical CNNs performed comparably to or better than complex hybrid models when using anatomically focused preprocessing and optimization.

Conclusions:

Anatomically constrained preprocessing via breast-mask segmentation significantly enhances the performance and stability of deep learning models for CEM breast lesion classification.
Input data quality and training optimization are critical for achieving clinically relevant AI performance, often more so than architectural complexity.
These findings support the development of more reliable AI-assisted decision support tools for CEM workflows.