ROBUSTNESS OF DEEP LEARNING MODELS TO DOMAIN SHIFT IN AUTOMATIC MELANOMA DETECTION ON DERMATOSCOPIC IMAGES

Abstract

Introduction. The article addresses the problem of automated melanoma detection in
dermoscopic images using deep learning methods under conditions of limited data availability, which
may contribute to the early diagnosis of this life-threatening disease. The main objective is to explore
ways of improving the generalization ability of deep learning models in classifying data from different
sources. The study utilised two sets of dermatoscopic images from the ISIC Archive in various
combinations. A series of experiments was conducted applying transfer learning with CNNs and the
Vision Transformer (ViT) architecture, a custom convolutional model, and model ensembles.
Particular attention is given to the impact of image preprocessing (removal of black borders,
augmentations) on classification quality, the use of data from other sources only for the minority class
and for all classes, as well as the generalization capability of models on cross-domain data. The
results demonstrated the advantage of transfer learning, while the custom model showed lower
accuracy. Among convolutional neural networks, the EfficientNet architecture achieved the highest
classification performance, while Vision Transformer models demonstrated strong potential,
outperforming most CNN architectures. The study revealed that combining data from different sources
may not improve, but rather decrease classification performance due to the model learning domain
differences instead of medical features. This effect was especially evident when supplementing only the
minority class from external datasets to address class imbalance during training. By removing noise in
the form of black borders characteristic of one dataset, the accuracy of disease recognition was
improved. The results of the study indicate the necessity of thorough data preparation and the
application of specialized strategies for cross-domain training to ensure practical applicability.
Purpose. The aim of the article is to investigate the impact of combining cross-domain data and
image preprocessing on the quality of melanoma classification using deep learning methods.
Results. The study reveals that combining images from different sources often leads not to
improvement but to a decline in performance. Mixing cross-domain data during model training can
introduce systematic classification bias, where the algorithm begins to associate image features not
with disease characteristics but with the domain from which the data originate. As a result, the model
learns to distinguish between data sources rather than actual medical indicators of pathology, which
negatively affects its generalization ability and diagnostic accuracy. Despite the increased overall
volume of data, differences in statistical characteristics between domains (such as imaging types,
lighting conditions, image formats, and class balance) complicate model training and may reduce its
ability to discriminate classes within individual domains. This is reflected in a decline in accuracy and
recall, even when AUC value remains high, indicating that the model cannot effectively separate
classes at realistic thresholds–a critical limitation for clinical use. Conclusion. From a practical perspective, the results of this work highlight the importance of
thorough data preparation and the development of adaptive strategies for cross-domain training.
Despite these challenges, developing models capable of reliably operating on heterogeneous data
remains a promising direction, as this approach most closely reflects the conditions of real-world
medical practice.