Scholarlink

Breast cancer is the most prevalent cancer among women. As per data from the Philippine Statistics Office and the Department of Health, roughly 3% of Filipino women are likely to receive a breast cancer diagnosis during their lifetime. Mammography has proven effective in identifying even early-stage breast cancers, making annual screenings crucial, particularly for women over 40. This study attempted to construct a hybrid model that segments cancerous cells in breast mammogram images. The proposed hybrid model utilized Res- Net152 on the Encoder block of Nested UNet (UNet++) model for semantic segmentation of cancerous cells in mammography. Intersection Over Union, DICE, and Binary Cross Entropy with Logits loss were used as indicators to measure the segmentation performance of the hybrid model. Custom Pre-Processing, Data Augmentation, Hyperparameters, and Fine Tuning were utilized in the study. The researchers achieved the mean and median detection score of IOU (67.57%) and DICE (68.40%), experimental results showed that the proposed model did great with limited dataset and did not overfit or underfit according to the study of Koehrsen.

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

Automated Semantic Segmentation of Cancerous Cells in Mammogram Images Using CNN Techniques

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Isidro P. Ampig,
  • Zuriel Jett M. Leung,
  • Vera Kim S. Tequin

摘要

Breast cancer is the most prevalent cancer among women. As per data from the Philippine Statistics Office and the Department of Health, roughly 3% of Filipino women are likely to receive a breast cancer diagnosis during their lifetime. Mammography has proven effective in identifying even early-stage breast cancers, making annual screenings crucial, particularly for women over 40. This study attempted to construct a hybrid model that segments cancerous cells in breast mammogram images. The proposed hybrid model utilized Res- Net152 on the Encoder block of Nested UNet (UNet++) model for semantic segmentation of cancerous cells in mammography. Intersection Over Union, DICE, and Binary Cross Entropy with Logits loss were used as indicators to measure the segmentation performance of the hybrid model. Custom Pre-Processing, Data Augmentation, Hyperparameters, and Fine Tuning were utilized in the study. The researchers achieved the mean and median detection score of IOU (67.57%) and DICE (68.40%), experimental results showed that the proposed model did great with limited dataset and did not overfit or underfit according to the study of Koehrsen.