DERMOSAN: Development of an interactive software tool in Streamlit for dermatological diagnosis assisted by convolutional neural networks (CNNs)

<ns3:p>Abstract* Background Artificial intelligence (AI)-assisted dermatological diagnosis has become increasingly important due to its ability to support clinical decision-making and reduce diagnostic variability among specialists. This paper presents DERMOSAN, an interactive software tool developed in Streamlit, based on a convolutional neural network (CNN), designed for the automatic classification of dermatological diseases using clinical images. Methods A public Kaggle dataset with 27,153 dermatological images distributed across ten clinical categories was used. The model was built on the ResNet152 architecture with transfer learning, implemented in TensorFlow and Keras. The images were divided into training (80%), validation (10%), and test (10%) subsets, applying class weighting to mitigate data imbalance. The tool integrates a Streamlit-based graphical interface that enables real-time image upload, local processing, and probabilistic diagnosis visualization. Results The model achieved 95% accuracy in training, 83% in validation, and 92.3% in testing. The confusion matrix showed better performance in the most representative classes. The developed interface allows for automatic diagnosis along with confidence scores and complementary clinical suggestions, facilitating quick and reproducible visual interpretation. Conclusions DERMOSAN is a stage of software development aimed at integrating deep learning models into accessible clinical environments. Although the system does not include clinical validation, it represents a step toward the creation of reproducible, open-source tools for AI-assisted diagnosis.</ns3:p>