Scholarlink

<p>Histopathological diagnosis of cutaneous drug eruptions (CDEs) presents a challenge in dermatopathology due to morphological overlap and high inter-observer variability. This retrospective study systematically compares the diagnostic performance of modern convolutional neural networks (ConvNeXt, ResNet 152, EfficientNet), hierarchical vision transformers (Swin Transformer), and traditional machine learning algorithms (HGBT, XGBoost) on whole-slide images (WSIs) of CDEs and validates decisions using explainable artificial intelligence (XAI). We analyzed 443 WSIs collected between 2014 and 2024, encompassing three diagnostic classes: CDE (<i>n</i> = 194), Atopic Dermatitis (<i>n</i> = 187), and histologically normal skin controls (<i>n</i> = 62), thereby better reflecting real-world clinical diagnostic challenges. Following controlled data augmentation to 600 images (200 per class), all models were evaluated on a held-out 20% test set (<i>n</i> = 120) using accuracy, sensitivity, specificity, precision, and F1-score across all three classes. Each WSI was processed as a single resized (512 × 512 pixels) sample; no sub-patch tessellation or patch-level aggregation was performed. Gradient-weighted Class Activation Mapping (Grad-CAM) was applied to visualize the histological regions influencing model decisions. ConvNeXt emerged as the highest-performing model, achieving the highest accuracy rate (88.98%) and demonstrating strong multi-class differential diagnosis capability. ResNet 152 and Hybrid Feature Fusion models also achieved highly competitive results (86.44% accuracy). ConvNeXt demonstrated a statistically significant advantage over traditional methods and legacy architectures (<i>p</i> &lt; 0.05). Histogram-Based Gradient Enhancement Classification Tree (HGBT) and XGBoost were effective among traditional models, with accuracies of 77.97% and 77.12%, respectively. Grad-CAM confirmed that models focused on clinically meaningful histological features, including basal vacuolar degeneration and eosinophilic infiltration. These findings suggest that advanced deep learning architectures, particularly ConvNeXt, may assist pathologists as a decision-support and triage tool in multi-class CDE diagnosis; ensemble methods offer a viable alternative in resource-constrained settings. Pending external validation, these XAI-supported models have the potential to provide transparent, reproducible diagnostic assistance in dermatopathology.</p>

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

Deep Learning vs. Traditional Machine Learning in the Diagnosis of Cutaneous Drug Eruptions: A Comparative Study with Explainable AI on Whole-Slide Images

  • 期刊论文

  • 发表日期：

    2026年6月18日

查看全文
  • Ayşe Türkmen Dedeoğlu,
  • Hüseyin Yanık,
  • Erdem Uysal,
  • Gizem Fikir,
  • Evren Değirmenci,
  • Ümit Türsen,
  • Yasemin Yuyucu Karabulut

摘要

Histopathological diagnosis of cutaneous drug eruptions (CDEs) presents a challenge in dermatopathology due to morphological overlap and high inter-observer variability. This retrospective study systematically compares the diagnostic performance of modern convolutional neural networks (ConvNeXt, ResNet 152, EfficientNet), hierarchical vision transformers (Swin Transformer), and traditional machine learning algorithms (HGBT, XGBoost) on whole-slide images (WSIs) of CDEs and validates decisions using explainable artificial intelligence (XAI). We analyzed 443 WSIs collected between 2014 and 2024, encompassing three diagnostic classes: CDE (n = 194), Atopic Dermatitis (n = 187), and histologically normal skin controls (n = 62), thereby better reflecting real-world clinical diagnostic challenges. Following controlled data augmentation to 600 images (200 per class), all models were evaluated on a held-out 20% test set (n = 120) using accuracy, sensitivity, specificity, precision, and F1-score across all three classes. Each WSI was processed as a single resized (512 × 512 pixels) sample; no sub-patch tessellation or patch-level aggregation was performed. Gradient-weighted Class Activation Mapping (Grad-CAM) was applied to visualize the histological regions influencing model decisions. ConvNeXt emerged as the highest-performing model, achieving the highest accuracy rate (88.98%) and demonstrating strong multi-class differential diagnosis capability. ResNet 152 and Hybrid Feature Fusion models also achieved highly competitive results (86.44% accuracy). ConvNeXt demonstrated a statistically significant advantage over traditional methods and legacy architectures (p < 0.05). Histogram-Based Gradient Enhancement Classification Tree (HGBT) and XGBoost were effective among traditional models, with accuracies of 77.97% and 77.12%, respectively. Grad-CAM confirmed that models focused on clinically meaningful histological features, including basal vacuolar degeneration and eosinophilic infiltration. These findings suggest that advanced deep learning architectures, particularly ConvNeXt, may assist pathologists as a decision-support and triage tool in multi-class CDE diagnosis; ensemble methods offer a viable alternative in resource-constrained settings. Pending external validation, these XAI-supported models have the potential to provide transparent, reproducible diagnostic assistance in dermatopathology.