Scholarlink

In this work, an improved deep learning algorithm for classifying the histopathology images of colorectal cancer is presented. By using state-of-the-art visual foundation models like Contrastive Language-Image Pretraining (CLIP), Distillation of Knowledge with No labels (DINOv2) and Bidirectional Encoder representations from image Transformer (BEiT) model, the image embeddings are obtained. The image embeddings are given to a Support Vector Classifier (SVC) and the labels of the test data are predicted. The final predictions of the test labels are estimated by using hard voting strategy on the predictions of the three models. The proposed algorithm is tested on NCT-CRC-HE-100K dataset which contains 1,00,000 images in training set and 7180 in validation set distributed across 9 classes. Accuracy, Precision, Recall and F1 scores of 97.31, 97.52, 97.31 and 97.27%, respectively, are obtained on the dataset and the results are superior to a recent work on the same dataset.

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

Colorectal Cancer Histopathology Images Classification Using Visual Foundation Models

  • 章节

  • 发表日期：

    2026年

查看全文
  • Bhuma Chandra Mohan

摘要

In this work, an improved deep learning algorithm for classifying the histopathology images of colorectal cancer is presented. By using state-of-the-art visual foundation models like Contrastive Language-Image Pretraining (CLIP), Distillation of Knowledge with No labels (DINOv2) and Bidirectional Encoder representations from image Transformer (BEiT) model, the image embeddings are obtained. The image embeddings are given to a Support Vector Classifier (SVC) and the labels of the test data are predicted. The final predictions of the test labels are estimated by using hard voting strategy on the predictions of the three models. The proposed algorithm is tested on NCT-CRC-HE-100K dataset which contains 1,00,000 images in training set and 7180 in validation set distributed across 9 classes. Accuracy, Precision, Recall and F1 scores of 97.31, 97.52, 97.31 and 97.27%, respectively, are obtained on the dataset and the results are superior to a recent work on the same dataset.