Developing breast lesion detection algorithms for digital breast tomosynthesis: Leveraging false positive findings

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Radiology

Background:

• Digital breast tomosynthesis (DBT) reading times exceed those of 2D mammography due to imaging complexity.
• A computer-aided diagnosis (CAD) system for DBT can alleviate radiologist workload and reduce reading times.

Purpose of the Study:

• To develop DBT lesion detection algorithms using multi-depth convolutional models and non-biopsied samples.
• To investigate the utility of false positive (FP) findings from non-biopsied benign lesions as data augmentation for improving detection algorithms.

Main Methods:

• Synthesized 2.5D images by combining lesion slices with adjacent slices for improved z-direction continuity.
• Employed YOLOv5 as the base network, training a baseline model and fine-tuning it with augmented data including actionable FPs.
• Developed an ensemble model by combining medium and large-depth level detection models for enhanced lesion inferencing.

Main Results:

• The baseline model achieved 0.640 sensitivity at 2 FPs per image (2FPI).
• Augmenting with actionable FPs improved the model to 0.769 sensitivity at 2FPI (p=0.013).
• The ensemble model achieved 0.80 sensitivity at 2FPI on the validation set (p<0.001) and 0.743 on the test set.

Conclusions:

• Actionable false positive findings contain valuable information for enhancing lesion detection algorithms.
• The proposed ensemble detection model, incorporating multi-depth levels, significantly improves lesion detection performance in DBT.