Avoiding Tissue Overlap in 2D Images: Single-Slice DBT Classification Using Convolutional Neural Networks.

Area of Science:

• Oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Breast cancer is a leading global diagnosis, making early detection critical.
• Mammography, a standard screening tool, suffers from tissue overlapping, obscuring lesions or creating false positives.
• Digital breast tomosynthesis (DBT) offers improved visualization by reconstructing 3D images from multiple angular views.

Purpose of the Study:

• To develop and evaluate a deep learning model for classifying breast lesions as benign or malignant using single DBT slices.
• To address the limitations of mammography by leveraging the non-overlapping nature of individual DBT slices.

Main Methods:

• A deep learning model, adapted from existing work, was trained on 2772 augmented DBT images (9 slices per volume from 77 DBT scans).
• Data augmentation techniques including rotation, translation, and mirroring were applied to enhance data variability.
• The model was evaluated on a separate test set derived from the same augmentation process.

Main Results:

• The model achieved a high accuracy of 93.2% on the testing set.
• Performance metrics included 92% sensitivity, 94% specificity, 94% precision, 94% F1-score, and 0.86 Cohen's kappa.
• These results indicate strong performance in distinguishing benign from malignant lesions.

Conclusions:

• Single-slice DBT analysis using this deep learning model demonstrates performance comparable to state-of-the-art methods.
• The findings suggest potential for AI-powered DBT slice analysis to rival mammography in diagnostic accuracy.
• Future research with larger datasets and advanced techniques like transfer learning could further enhance diagnostic capabilities.