Sign in Subscribe
Artificial neural networks

Robustness of Deep Learning-Convolutional Neural Networks in Dermoscopy: A Cross-Sectional Study

Strategic Insights

— 1 min read

A new study has been conducted to evaluate the robustness of deep learning-convolutional neural networks (DL-CNNs) in dermoscopy when challenged with close-up images. The study aimed to assess the diagnostic accuracy of DL-CNNs trained on dermoscopic images when applied to close-up images of skin lesions. The researchers used a cross-sectional study design, where a DL-CNN, named Moleanalyzer pro, was trained on 129,487 dermoscopic images and tested on 350 skin lesions, each imaged both clinically and dermoscopically. The study's primary outcomes included sensitivity, specificity, and the receiver operating characteristic-area under the curve (ROC-AUC).

Key Takeaways:

  • The DL-CNN, Moleanalyzer pro, achieved a sensitivity of 88.2% (95% CI: 82.9%-92.0%) and specificity of 69.0% (61.4%-75.8%) for dermoscopic images.
  • When challenged with close-up images, the sensitivity of the DL-CNN decreased to 60.5% (53.5%-67.1%, p < 0.001).
  • Fine-tuning of DL-CNNs on clinical images may enhance real-world diagnostic accuracy.
  • The study used a cross-sectional design with 350 skin lesions, each imaged both clinically and dermoscopically.
  • The DL-CNN, Moleanalyzer pro, was trained on 129,487 dermoscopic images.
  • The study's results suggest that DL-CNNs may not perform optimally when applied to close-up images of skin lesions.

Statistics:

  • Sensitivity of the DL-CNN for dermoscopic images: 88.2% (95% CI: 82.9%-92.0%).
  • Sensitivity of the DL-CNN for close-up images: 60.5% (53.5%-67.1%, p < 0.001).
  • Specificity of the DL-CNN for dermoscopic images: 69.0% (61.4%-75.8%).
  • Receiver operating characteristic-area under the curve (ROC-AUC) for dermoscopic images: 0.866 (0.860-0.873).
  • Total number of skin lesions used in the study: 350.
  • Total number of dermoscopic images used to train the DL-CNN: 129,487.

Sources:

  • Robustness of a convolutional neural network trained on dermoscopic images and challenged with close-up images. JDDG, 2025.
  • Wiley, 111 River St, Hoboken 07030-5774, NJ, USA.
  • onlinelibrary.wiley.com/journal/10.1111/(ISSN)1610-0387
  • Julia Katharina Winkler, Dept. of Dermatology, University medical Center Heidelberg, Heidelberg, Germany.
  • Anastasia Sophie Vollmer, Katharina Susanne Kommoss, Wilhelm Stolz, Albert Rosenberger, Alexander Enk, and Holger Andreas Haenssle.

Read more