Selecting the most appropriate pre-trained CNN architecture for each imaging modality

Determine which pre-trained convolutional neural network architecture is most appropriate for melanoma detection across dermatoscopic and histopathological imaging modalities, accounting for differences in image characteristics and acquisition conditions.

Background

The paper compares ResNet50, VGG16, VGG19, MobileNet, and InceptionV3 under a common experimental protocol on the HAM10000, ISIC 2018, and CR-AI4SkIN datasets. The results show that architecture performance varies across dermatoscopic and histopathological images, indicating that no single architecture can be assumed to be optimal independently of the imaging modality.

The unresolved issue is the broader selection problem: identifying the architecture that is most suitable for a particular melanoma-imaging context while achieving robust and reliable performance across differing data characteristics. The paper contributes comparative evidence but does not establish a general selection rule or universally optimal architecture.