Skin Cancer Classification: A Comparison of CNN-Backbones for Feature-Extraction

In order to classify histopathology images into 11 classes of cancerous and non-cancerous skin diseases, the pharmaceutical bioinformatics research group at University Freiburg implemented a pipeline based on the TransMIL model [1]. To improve the general performance, we compared four different CNNs for feature extraction in the TransMIL preprocessing pipeline CLAM [2]. Comprehensive evaluations, including detailed analyses of loss, accuracy, F1 score, and attention focus, indicate that ConvNextV2 [3] falls short in comparison to ResNet50 [4], DenseNet201 [5], and EfficientNet [6], which demonstrate nearly equal performance across classes and achieve higher average accuracy. Surprisingly, despite its size, EfficientNet showed slightly better results compared to DenseNet201 and ResNet50. Moreover, the present pipeline exhibits uneven performance across various disease classes, with particular difficulty in distinguishing between the three classes, likely due to the inherent complexities associated with these categories. Despite these challenges, the EfficientNet model remains the most balanced among all those evaluated.