Automatic Bone Metastasis Classification: An in-depth Comparison of CNN and Transformer Architectures

Automatic classification of bone metastases is a major challenge and is receiving increasing attention from the research community. One of the major challenges is the accurate classification of medical images, especially the distinction between benign and malignant images, which can greatly help physicians in decision-making. Recently, several deep-learning techniques have been proposed for medical image classification. Their performance, however, is influenced by both the dataset and the imaging modality. In this work, we investigate the performance of several state-of-the-art CNN architectures, namely InceptionV3, EfficientNet, ResNext50, and DenseNet161, as well as Transformer architectures, namely ViT and DeiT. We trained and tested these algorithms on a large dataset consisting of CT-scan images. The Transformer algorithms were found to be superior to CNN algorithms in detecting bone metastases. In particular, ViT Tiny achieved the best performance in terms of accuracy and F1-score as compared to other architectures.