Enhancing Automated Melanoma Detection: Comparative Analysis of CNN Architectures Using the SIIM-ISIC 2020 Kaggle Dataset

Abstract

This investigation explores the combination of AI and behavioural analytics for melanoma detection, integrating deep learning model assessment with the statistical analysis of factors influencing user adoption. The work compares the diagnostic efficiency and interpretability of three state-of-the-art convolutional neural networks—EfficientNetB0, ResNet50, and MobileNetV2—using the SIIM-ISIC 2020 dataset. The preprocessing of data was performed by image normalisation, resizing, and augmentation of the images to prevent overfitting and improve generalisation. Transfer learning was done using ImageNet weights, with models trained using the Adam optimizer and categorical crossentropy loss. The performance of the models was measured in terms of accuracy, precision, recall, F1-score, and AUC-ROC, whereas Grad-CAM visualisation was used to provide information about the localisation of the lesion and understanding. The computational results were supported by a quantitative study performed using SPSS, which focused on the healthcare professionals’ perceptions, attitudes, and willingness to adopt AI-based melanoma detection systems. The survey data were processed by descriptive statistics of 100 respondents, correlation analysis, and hypothesis testing through Spearman’s Rank Correlation and regression models. These tests explored the impact of performance
expectancy, effort expectancy, social influence, facilitating conditions, and ethical trust on user acceptance. Findings disclosed significant positive correlations between perceived usefulness, trust, and adoption intention that were consistent with the human-centric dimension of AI integration in the clinical context. In general, the EfficientNetB0 model was able to obtain the top AUC-ROC score of 0.5523; however, all of the models showed low sensitivity towards melanoma because of the imbalance of the dataset. The results emphasise that, on the technical side, performance alone is not enough for successful implementation; proper training, sophisticated loss functions (like focal loss), and ethical transparency are still necessary. The deep learning and SPSS studies, taken together, suggest that the dermatology AI intervention will be effective only if the algorithmic issues are resolved and there is social, ethical, and professional agreement with the clinical stakeholders.