Related Experiment Video
Updated: Sep 19, 2025
13:44
Detection of Architectural Distortion in Prior Mammograms via Analysis of Oriented Patterns
Published on: August 30, 2013
43.0K
Comparative analysis of semantic-segmentation models for screen film mammograms
Jyoti Rani1, Jaswinder Singh2, Jitendra Virmani3
1GZSCCET, MRSPTU, Punjab, India.
Computers in Biology and Medicine
|June 6, 2025
Summary
Accurate mammographic mass segmentation is crucial for diagnosis. ResNet50 demonstrated superior performance among ten deep learning models, proving effective for segmenting challenging breast masses in mammography screening.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Computer Vision
Background:
- Accurate segmentation of mammographic masses is critical for radiologists to differentiate between benign and malignant cases based on shape characteristics.
- Deep learning segmentation algorithms have gained prominence for their application in medical image analysis tasks.
Purpose of the Study:
- To conduct a rigorous performance analysis of ten semantic segmentation models for mammographic mass segmentation.
- To identify the optimal deep learning model for segmenting challenging breast masses, including those with dense backgrounds or co-occurring micro-calcifications.
Main Methods:
- Evaluation of ten semantic segmentation models (VGG16/VGG19, U-Net, ResNet18/ResNet50/ShuffleNet/XceptionNet/InceptionV2/MobileNetV2, hybrid U-Net) using 518 mammographic images from the DDSM dataset.
- Quantitative assessment using Jaccard Index (JI) and F1 scores.
- Subjective analysis of segmented images by radiologists focusing on size, margins, and shape characteristics.
Main Results:
- Dilated convolution DAG models, specifically ResNet50, ShuffleNet, and MobileNetV2, outperformed other models.
- ResNet50 achieved cumulative JI and F1 scores of 0.87 and 0.92, respectively.
- ResNet50 was identified as the optimal model for segmenting difficult-to-delineate masses, including those with dense backgrounds and co-existing micro-calcifications.
Conclusions:
- ResNet50 is the most effective model for segmenting mammographic masses, particularly complex cases.
- The study indicates the potential for ResNet50 to be integrated into routine clinical workflows for mammographic mass segmentation.
- Deep learning models, especially ResNet50, offer significant advancements in improving the accuracy and efficiency of mammographic analysis.
