Deep learning methods for lesion detection on mammography images: a comparative analysis.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Breast cancer diagnosis relies on mammography, but lesion segmentation is difficult due to low contrast, variable sizes, and small lesion detection challenges.
• Deep learning (DL) methods have emerged as powerful tools for medical image segmentation, offering potential solutions to these segmentation difficulties.

Purpose of the Study:

• To benchmark and compare the performance of six state-of-the-art deep learning architectures for breast lesion segmentation in mammography images.
• To evaluate segmentation methods using key metrics including Dice coefficient, Jaccard index, accuracy, recall, specificity, and precision.

Main Methods:

• Six DL architectures (U-Net, UNETR, DynUNet, SegResNetVAE, RF-Net, MDA-Net) were trained on 1692 mammograms from the CBIS-DDSM dataset.
• A combination of cross-entropy and Dice loss functions was used for training.
• Performance was assessed using Dice coefficient, Jaccard index, accuracy, recall, specificity, and precision.

Main Results:

• All evaluated DL networks achieved Dice scores exceeding 86%.
• MDA-Net and DynUnet demonstrated superior performance, achieving Dice scores of 90.25% and 89.67%, respectively.
• MDA-Net and DynUnet also reported high accuracy rates of 93.48% and 93.03%, respectively.

Conclusions:

• Deep learning strategies, particularly MDA-Net and DynUnet, are highly effective for breast lesion segmentation in mammography.
• This comparative study provides valuable insights into the current capabilities of DL for improving breast cancer diagnosis through automated segmentation.