Scholarlink

Chronic Kidney Diseases (CKD), such as cysts, tumors, and stones, pose significant diagnostic challenges due to their visual similarity in CT scans. Manual interpretation is time-consuming and susceptible to variability, prompting the need for automated, accurate classification systems. This study compares the performance of two feature extraction approaches for multiclass kidney disease classification from CT images: (i) deep transfer learning using EfficientNet-B0 and (ii) handcrafted Histogram of Oriented Gradients (HOG) features. Both were classified using a Random Forest (RF) model under a standardized evaluation framework. A balanced dataset of 4,880 CT images was labeled into four categories—Normal, Cyst, Tumor, and Stone. Features extracted via EfficientNet-B0 and HOG were independently passed to a Random Forest classifier. Performance was assessed using confusion matrices, accuracy, precision, recall, F1-score, and ROC-AUC. The EfficientNet + RF pipeline achieved 99.63% accuracy, 99.76% precision, 99.32% recall, and 99.54% F1-score, with AUC = 1.00 for all classes. In comparison, the HOG + RF model recorded 99.81% accuracy, 99.83% precision, 99.60% recall, and 99.71% F1-score, also with perfect AUC values. However, HOG exhibited slightly higher misclassification, especially in the Cyst class. While both models performed exceptionally, EfficientNet demonstrated improved differentiation between visually similar classes such as Cyst and Tumor, as reflected in lower misclassification rates in the confusion matrix.

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

A Comparative Analysis of Deep Transfer Learning and Traditional Handcrafted Feature Approaches for Kidney Disease Classification Using Random Forests

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Roseline Oluwaseun Ogundokun,
  • Pius Adewale Owolawi,
  • Etienne A. van Wyk

摘要

Chronic Kidney Diseases (CKD), such as cysts, tumors, and stones, pose significant diagnostic challenges due to their visual similarity in CT scans. Manual interpretation is time-consuming and susceptible to variability, prompting the need for automated, accurate classification systems. This study compares the performance of two feature extraction approaches for multiclass kidney disease classification from CT images: (i) deep transfer learning using EfficientNet-B0 and (ii) handcrafted Histogram of Oriented Gradients (HOG) features. Both were classified using a Random Forest (RF) model under a standardized evaluation framework. A balanced dataset of 4,880 CT images was labeled into four categories—Normal, Cyst, Tumor, and Stone. Features extracted via EfficientNet-B0 and HOG were independently passed to a Random Forest classifier. Performance was assessed using confusion matrices, accuracy, precision, recall, F1-score, and ROC-AUC. The EfficientNet + RF pipeline achieved 99.63% accuracy, 99.76% precision, 99.32% recall, and 99.54% F1-score, with AUC = 1.00 for all classes. In comparison, the HOG + RF model recorded 99.81% accuracy, 99.83% precision, 99.60% recall, and 99.71% F1-score, also with perfect AUC values. However, HOG exhibited slightly higher misclassification, especially in the Cyst class. While both models performed exceptionally, EfficientNet demonstrated improved differentiation between visually similar classes such as Cyst and Tumor, as reflected in lower misclassification rates in the confusion matrix.