Automatic Detection and Segmentation of Breast Cancer on MRI Using Mask R-CNN Trained on Non-Fat-Sat Images and...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Computer-aided diagnosis is crucial for interpreting breast MRI.
• Early and accurate detection of breast lesions improves patient outcomes.
• Deep learning offers advanced capabilities for image analysis.

Purpose of the Study:

• To evaluate the efficacy of a deep learning Mask Regional Convolutional Neural Network (R-CNN) for detecting lesions in breast MRI.
• To assess the performance of Mask R-CNN in localizing and segmenting suspicious breast lesions.
• To explore the potential of AI in developing automated breast MRI diagnostic systems.

Main Methods:

• Utilized two dynamic contrast-enhanced MRI (DCE-MRI) datasets for training and testing.
• Employed a Mask R-CNN with ResNet-101 backbone for lesion detection and segmentation.
• Evaluated performance using Dice Similarity Coefficient (DSC) and free-response receiver operating characteristic analysis.

Main Results:

• Mask R-CNN achieved high accuracy (0.86 training, 0.75 testing) and DSC (0.82 training, 0.79 testing).
• The model successfully identified 99.5% of lesions in the training set and 100% in the testing set.
• Minimizing false positives was achieved by incorporating precontrast and subtraction images.

Conclusions:

• Deep learning with Mask R-CNN is a feasible method for breast MRI lesion detection and segmentation.
• This AI approach can significantly aid in the localization and characterization of breast lesions.
• Integration with other AI algorithms can lead to fully automated breast MRI diagnostic systems.