Convolutional neural network for breast cancer diagnosis using diffuse optical tomography

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Breast cancer screening relies on accurate lesion detection.
• Diffuse optical tomography (DOT) offers potential for repeated measurements in mass screening.
• Computer-aided diagnosis (CAD) systems can enhance diagnostic accuracy.

Purpose of the Study:

• To develop and evaluate a convolutional neural network (CNN)-based CAD system for classifying breast mass lesions.
• To assess the system's performance on optical tomographic images from women with dense breasts.
• To improve classification performance using data augmentation.

Main Methods:

• A CNN was developed to classify breast mass lesions from 2D grayscale images derived from 3D DOT data.
• A dataset of 1260 images was created from 63 optical tomographic scans.
• Image preprocessing, normalization, and data augmentation techniques were applied.

Main Results:

• The initial CNN model achieved 0.80 specificity, 0.95 sensitivity, 90.2% accuracy, and 0.94 AUC.
• After data augmentation, performance improved to 0.88 sensitivity, 0.96 specificity, 93.3% accuracy, and 0.95 AUC.
• The system demonstrated robust classification of benign and malignant breast lesions.

Conclusions:

• The developed CNN-based CAD system shows high efficacy in classifying breast mass lesions from DOT images.
• Data augmentation significantly enhanced the system's diagnostic performance, particularly for imbalanced datasets.
• This AI-driven approach holds promise for improving the accuracy and efficiency of breast cancer screening.