Scholarlink

The differential diagnosis between breast masses belonging to probably benign, suspicious malignant and malignant classes is a challenge faced by radiologists primarily due to presence of overlapping shape and textural features given that these masses are often buried in various types of background tissue densities. Initially, 518 mammogram images from DDSM database are subjected to ResNet50 segmentation model to obtain segmented masses. The textural features and morphological features computed from the segmented masses are then subjected to ANFC-LH/PCA-SVM classification algorithms. Two types of CAD systems designs (1) single multi-class classification framework and (2) hierarchical classification framework with 02 binary classifiers have been implemented. It can be concluded that hybrid hierarchical framework with ANFC-LH at node 1 and PCA-SVM at node 2 provides reasonable accuracy of 75% and 70% considering varying breast densities and highly overlapping appearances of benign, suspicious malignant and malignant classes.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

ML Based Multi Class and Hierarchical Classification Frameworks for DDSM Images

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Jyoti Rani,
  • Jaswinder Singh,
  • Jitendra Virmani

摘要

The differential diagnosis between breast masses belonging to probably benign, suspicious malignant and malignant classes is a challenge faced by radiologists primarily due to presence of overlapping shape and textural features given that these masses are often buried in various types of background tissue densities. Initially, 518 mammogram images from DDSM database are subjected to ResNet50 segmentation model to obtain segmented masses. The textural features and morphological features computed from the segmented masses are then subjected to ANFC-LH/PCA-SVM classification algorithms. Two types of CAD systems designs (1) single multi-class classification framework and (2) hierarchical classification framework with 02 binary classifiers have been implemented. It can be concluded that hybrid hierarchical framework with ANFC-LH at node 1 and PCA-SVM at node 2 provides reasonable accuracy of 75% and 70% considering varying breast densities and highly overlapping appearances of benign, suspicious malignant and malignant classes.