Scholarlink

This study examines the comparative performance of Vision Transformers in classifying breast cancer images under various pre-processing conditions. The analysis was conducted on raw breast cancer mammograms, as well as images enhanced using Gamma correction and Contrast Limited Adaptive Histogram Equalization. Specifically, five dataset configurations were considered: (1) original images, (2) Gamma-corrected images only, (3) original images combined with Gamma-corrected versions, (4) Contrast Limited Adaptive Histogram Equalization-enhanced images only, (5) original images combined with Contrast Limited Adaptive Histogram Equalization-enhanced versions. Additionally, the analysis was stratified by lesion types and mammographic view, comparing model performance for calcification and mass lesions on Mediolateral Oblique and Craniocaudal projections. The results reveal distinct performance patterns in the vision transformer model across these preprocessing scenarios and imaging perspectives, highlighting the impact of image enhancement techniques and view types on model performance. This study provides valuable insights into the effectiveness of pre-processing methods in optimising the performance of advanced neural network architectures for breast cancer image classification. The model demonstrates strong performance across both datasets, achieving its highest results on MIAS dataset with an accuracy, sensitivity, and specificity of 0.99 and on DDSM dataset with an accuracy of 0.95, sensitivity of 0.96 and specificity of 0.95.

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

Fine-Tuned Vision Transformers for Mammographic Breast Cancer Classification: Evaluating the Impact of Preprocessing Techniques

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Bianca Iacob,
  • Laura Diosan

摘要

This study examines the comparative performance of Vision Transformers in classifying breast cancer images under various pre-processing conditions. The analysis was conducted on raw breast cancer mammograms, as well as images enhanced using Gamma correction and Contrast Limited Adaptive Histogram Equalization. Specifically, five dataset configurations were considered: (1) original images, (2) Gamma-corrected images only, (3) original images combined with Gamma-corrected versions, (4) Contrast Limited Adaptive Histogram Equalization-enhanced images only, (5) original images combined with Contrast Limited Adaptive Histogram Equalization-enhanced versions. Additionally, the analysis was stratified by lesion types and mammographic view, comparing model performance for calcification and mass lesions on Mediolateral Oblique and Craniocaudal projections. The results reveal distinct performance patterns in the vision transformer model across these preprocessing scenarios and imaging perspectives, highlighting the impact of image enhancement techniques and view types on model performance. This study provides valuable insights into the effectiveness of pre-processing methods in optimising the performance of advanced neural network architectures for breast cancer image classification. The model demonstrates strong performance across both datasets, achieving its highest results on MIAS dataset with an accuracy, sensitivity, and specificity of 0.99 and on DDSM dataset with an accuracy of 0.95, sensitivity of 0.96 and specificity of 0.95.