Performance Analysis of Convolutional Neural Network Architectures for Automated Melanoma Detection

Abstract

Deep-learning-based automated melanoma detection has shown good diagnostic capability, but its potential in clinical environment is not only based on the predictive power but also on the trust and acceptance of clinicians. This paper presents the convolutional neural network (CNN) based melanoma recognition with a survey-based measure of the adoption intentions of clinicians. Three pre-trained CNN models, efficientnetb0, ResNet50 and MobileNetV2 were tested in terms of the SIIM-ISIC 2020 dermoscopic image dataset on standardized preprocessing, data augmentation and training conditions. The accuracy of EfficientNetB0 was 0.9817, the ResNet50 was at 0.9804, and the MobileNetV2 was 0.9829 which is indicative of good classification. A survey based on the Unified Theory of Acceptance and Use of Technology (UTAUT) and trust-ethics concepts and developed on the survey of 100 clinicians was carried out to supplement the technical assessment. The rank-order analysis of correlation of Spearman (n = 100) showed that acceptance of AI-based melanoma detection positively correlated with performance expectancy (r = 0.540), effort expectancy (r = 0.591), social influence (r = 0.571), facilitating conditions (r = 0.636), trust (r = 0.537), and ethical/privacy concerns (r = 0.544) and with p ≤ 0.01. These statistically significant associations indicate that technological and trust-related factors are determinants of clinicians' intentions to adopt them. The results indicate that high technical performance should be combined with trust, ethical transparency, and clinician acceptance to facilitate the real-world implementation of AI-based diagnostic tools.

Ähnliche Arbeiten