Convolutional neural network for automated mass segmentation in mammography

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Mammogram (MG) image analysis for lesion segmentation is challenging.
• Deep learning (DL) methods offer advanced solutions but require precise models.
• Developing accurate models for mass lesion segmentation in MG images is crucial for early cancer detection.

Purpose of the Study:

• To develop a novel deep learning model for precise segmentation of mass lesions in mammogram images.
• To improve the accuracy and efficiency of automated lesion detection in mammography.
• To enhance radiologists' ability to differentiate between benign and malignant lesions.

Main Methods:

• A modified U-Net architecture, termed Vanilla U-Net, was developed for semantic segmentation.
• The model combines low-level and high-level features to extract contextual information.
• Training was performed on multiple large-scale public (CBIS-DDSM, BCDR-01, INbreast) and private mammogram databases.
• Data augmentation techniques were employed to enhance model robustness and performance.

Main Results:

• The Vanilla U-Net model demonstrated superior performance compared to state-of-the-art models like Faster R-CNN, FCN, SegNet, Dilated-Net, and original U-Net.
• Achieved a mean test accuracy of 92.6% on digitized film-based and fully digitized MG images.
• Obtained a mean Dice coefficient index (DI) of 0.951 and a mean Intersection over Union (IOU) of 0.909.
• Data augmentation significantly improved mean DI from 0.922 to 0.951 and mean IOU from 0.856 to 0.909.

Conclusions:

• The proposed Vanilla U-Net model enables precise pixel-wise segmentation of masses in mammograms.
• The model's ability to capture multi-scale spatial context aids in differentiating lesion types.
• Transfer learning, data augmentation, and architectural modifications led to superior performance over conventional and other DL models.