Scholarlink

Image segmentation is the challenging task of predicting a semantic class value for every pixel in an image. This work introduces two deep learning model architectures, Padded Convolution (PC) U-Net and Half Cut (HC) W-Net, that primarily solve the task of image segmentation on a collection of colonoscopy images designed to detect polyps for early detection of cancer. PC U-Net uses a symmetrical structure with padded convolutions, ensuring consistent pixel alignment between input and output images while preserving border information. HC W-Net extends this by employing a W-shaped architectural design and incorporating skip connections to retain information. We show that both models aim to improve the automatic detection of polyps, which is critical for early diagnosis and prevention of colorectal cancer. The best mIoU achieved on train dataset in PC U-Net and HC W-Net is 0.6553 and 0.8466, respectively. Moreover, on test dataset, the best mIoU obtained in PC U-Net and HC W-Net is 0.5287 and 0.5843, respectively.

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

PC U-Net and HC W-Net: Convolutional Neural Network Architectures for Image Segmentation

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Hasan Shakir,
  • Aysha Athar Siddiqui,
  • M. Nizamuddin

摘要

Image segmentation is the challenging task of predicting a semantic class value for every pixel in an image. This work introduces two deep learning model architectures, Padded Convolution (PC) U-Net and Half Cut (HC) W-Net, that primarily solve the task of image segmentation on a collection of colonoscopy images designed to detect polyps for early detection of cancer. PC U-Net uses a symmetrical structure with padded convolutions, ensuring consistent pixel alignment between input and output images while preserving border information. HC W-Net extends this by employing a W-shaped architectural design and incorporating skip connections to retain information. We show that both models aim to improve the automatic detection of polyps, which is critical for early diagnosis and prevention of colorectal cancer. The best mIoU achieved on train dataset in PC U-Net and HC W-Net is 0.6553 and 0.8466, respectively. Moreover, on test dataset, the best mIoU obtained in PC U-Net and HC W-Net is 0.5287 and 0.5843, respectively.