A deep learning classifier for digital breast tomosynthesis

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Healthcare
• Machine Learning for Medical Diagnosis

Background:

• Digital breast tomosynthesis (DBT) is an advanced imaging technique for breast cancer screening.
• Accurate detection of mass lesions in DBT images is crucial for early diagnosis and treatment.
• Computerized detection systems can assist radiologists in interpreting complex DBT data.

Purpose of the Study:

• To develop and evaluate a deep convolutional neural network (DCNN) for automatic classification of mass lesions in DBT images.
• To compare the performance of different DCNN architectures for this task.
• To investigate the potential of Grad-CAM for lesion localization within DBT slices.

Main Methods:

• Three DCNN architectures (AlexNet, VGG19, and a custom DBT-DCNN) were trained and evaluated on DBT slice images.
• Image preprocessing included normalization, background correction, and data augmentation.
• Performance was assessed using accuracy, sensitivity, AUC, and Grad-CAM for lesion visualization.

Main Results:

• The custom DBT-DCNN achieved high performance with 90% accuracy and 96% sensitivity.
• Area Under the Curve (AUC) was 0.89 ± 0.04, with cross-validation accuracy of 94.0% ± 0.2%.
• Grad-CAM maps effectively highlighted suspicious regions within the DBT images, indicating lesion locations.

Conclusions:

• A deep learning-based framework (DBT-DCNN) was successfully developed for classifying mass lesions in DBT images.
• The system demonstrates potential for aiding radiologists in breast cancer detection.
• Grad-CAM can be applied to identify lesion positions, enhancing the interpretability of the AI model.