https://atlas.rsna.org/schemas/2025-11/model.json

A MULTI RECONSTRUCTION STUDY OF BREAST DENSITY ESTIMATION USING DEEP LEARNING

https://arxiv.org/pdf/2202.08238

This paper presents a deep learning approach for breast density estimation using a multi-reconstruction strategy, combining various mammography acquisition protocols to improve model generalizability and performance.

The model uses transfer learning with Inception-V3, pre-trained on the ImageNet dataset. The last layers of the original Inception-V3 model were replaced with a fully connected layer containing 1024 nodes, followed by four output nodes with softmax activation. Training was performed using Keras with TensorFlow-1.10, an initial learning rate of 0.001 on a stochastic gradient descent optimizer with a batch size of 8, and terminated after 50 epochs. Traditional augmentation routines (random rotation, horizontal and vertical flipping, random crops, and translation) were applied.

Increased accuracy of breast density measurement beyond the quartile BI-RADS schema, aiding in recognizing individuals predisposed to breast cancer and improving real-world deployment and automatic reporting by being more generalizable across different imaging protocols.

Breast cancer screening, mammogram interpretation.

Automated breast density classification.

Estimation of breast density categories (A, B, C, D) from mammographic images for breast cancer risk assessment.

Full Field Digital Mammography (FFDM), Hologic’s 2D Intelligent view, C-view, 2D tomographic projections, XCCM (exaggerated cranio-caudal medial), XCCL (exaggerated cranio-caudal lateral), Latero medial, and Medio-lateral mammography views.

The model should be used with a variety of mammography acquisition protocols, including FFDM, C-View, Tomosynthesis Projection, Intelligent 2D, XCCL, XCCM, Latero medial, and Medio-lateral images. It is designed to perform comparably or better than models trained on single acquisition protocols. Continuous learning by adding more data as it is acquired is recommended for further improvement and generalizability.

Previous deep learning models for breast density estimation often rely on single acquisition protocols, limiting dataset size and generalizability, and performing poorly on images from different protocols. This study aims to mitigate these limitations by using a multi-reconstruction approach.

Researchers should combine images from different acquisition protocols to produce more generalizable neural network models capable of continuous learning as more data is collected.

The model's architecture (Inception-V3 with specific modifications), training framework (Keras/TensorFlow), optimizer (stochastic gradient descent), learning rate (0.001), batch size (8), epochs (50), and augmentation techniques are detailed, providing a basis for reproducibility.