Scholarlink

The classification of calcifications in mammograms is a critical task in the early detection of breast cancer, yet the performance of deep learning models is highly dependent on the quality of input images. This paper presents a study on the effects of various image preprocessing techniques on the performance of four different Convolutional Neural Network (CNN) architectures for the morphological classification of calcifications. The study is based on the public CBIS-DDSM dataset considering ROIs corresponding to the morphologies CLUSTERED and SEGMENTAL. Twelve transformations, including Contrast Limited Adaptive Histogram Equalization (CLAHE) with varying parameters, gamma correction, and the addition and subsequent filtering of salt-and-pepper noise are considered. The results demonstrate that the impact of preprocessing is architecture dependent. CLAHE proved to be the most effective technique for enhancing performance, achieving an accuracy improvement of up to 12.61% with MobileNetV2 compared to the original images. Similar enhancements were observed for ResNet50V2 (7.60% improvement) and VGG16 (2.10% improvement). Furthermore, the results confirm that filtering noisy images systematically improves classification accuracy in all tested scenarios. These findings highlight the critical importance of a co-designed preprocessing and network selection strategy to optimize the performance of automated diagnostic systems.

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

Evaluating Preprocessing Techniques for Calcification Morphology Classification in Mammograms Based on Convolutional Neural Networks

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Agustín Amalfitano,
  • Juan I. Iturriaga,
  • María D. Perez-Godoy,
  • Diego Sebastian Comas

摘要

The classification of calcifications in mammograms is a critical task in the early detection of breast cancer, yet the performance of deep learning models is highly dependent on the quality of input images. This paper presents a study on the effects of various image preprocessing techniques on the performance of four different Convolutional Neural Network (CNN) architectures for the morphological classification of calcifications. The study is based on the public CBIS-DDSM dataset considering ROIs corresponding to the morphologies CLUSTERED and SEGMENTAL. Twelve transformations, including Contrast Limited Adaptive Histogram Equalization (CLAHE) with varying parameters, gamma correction, and the addition and subsequent filtering of salt-and-pepper noise are considered. The results demonstrate that the impact of preprocessing is architecture dependent. CLAHE proved to be the most effective technique for enhancing performance, achieving an accuracy improvement of up to 12.61% with MobileNetV2 compared to the original images. Similar enhancements were observed for ResNet50V2 (7.60% improvement) and VGG16 (2.10% improvement). Furthermore, the results confirm that filtering noisy images systematically improves classification accuracy in all tested scenarios. These findings highlight the critical importance of a co-designed preprocessing and network selection strategy to optimize the performance of automated diagnostic systems.