Scholarlink

Diagnosing progressive skin conditions such as eczema and melanoma is difficult due to overlapping visual features and the scarcity of longitudinal data. Traditional single-image models cannot capture temporal dynamics that are critical for accurate diagnosis. In this study, we present a proof-of-concept dual framework that combines ConvLSTM-based temporal convolutional networks and TimeSformer-based spatio-temporal transformers, trained on week-by-week synthetic progression sequences generated via CycleGAN. Using 5-frame timelines synthesized from the DermNet dataset of 19,500 images, ConvLSTM achieved 98% accuracy, 0.97 F1-score, and 0.98 recall on binary classification tasks for eczema and melanoma, while multiclass classification across 23 skin diseases reached 78%. TimeSformer obtained 90% accuracy and 0.89 F1-score with temporal attention. Grad-CAM visualizations provided model interpretability. While limited by reliance on synthetic data, this framework highlights the potential of temporal modeling for skin disease progression and lays the foundation for future validation with real-world longitudinal datasets.

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

Explainable Skin Disease Diagnosis and Temporal Progression Modeling Using GAN-Generated Synthetic Data and Spatio-Temporal Attention Networks

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Abdullah Al Noman,
  • Chanda Rani Debi,
  • Sabrina Yesmin,
  • Mehedi Hassan Imon

摘要

Diagnosing progressive skin conditions such as eczema and melanoma is difficult due to overlapping visual features and the scarcity of longitudinal data. Traditional single-image models cannot capture temporal dynamics that are critical for accurate diagnosis. In this study, we present a proof-of-concept dual framework that combines ConvLSTM-based temporal convolutional networks and TimeSformer-based spatio-temporal transformers, trained on week-by-week synthetic progression sequences generated via CycleGAN. Using 5-frame timelines synthesized from the DermNet dataset of 19,500 images, ConvLSTM achieved 98% accuracy, 0.97 F1-score, and 0.98 recall on binary classification tasks for eczema and melanoma, while multiclass classification across 23 skin diseases reached 78%. TimeSformer obtained 90% accuracy and 0.89 F1-score with temporal attention. Grad-CAM visualizations provided model interpretability. While limited by reliance on synthetic data, this framework highlights the potential of temporal modeling for skin disease progression and lays the foundation for future validation with real-world longitudinal datasets.