Vision Transformer for Ovarian Tumor Classification: A Comparative Study with CNNs on Ultrasound Imaging

Ovarian cancer continues to be one of the most deadly gynaecological cancers worldwide. Early and accurate diagnosis is critically important. This study systematically compares Vision Transformer (ViT) architecture against established CNN models for ovarian tumor classification using the Multi-Modality Ovarian Tumor Ultrasound (MMOTU) dataset. We implemented ResNet50, Vision Transformer (ViT), EfficientNet-B0, EfficientNetV2-m, and DenseNet-121 models with a new three-class classification approach that groups the original eight tumor categories into clinically relevant classes: Benign, Malignant, and Normal. Our methodology incorporates stratified 5-fold cross-validation to address the limitation of single train-test splits in previous studies, comprehensive data augmentation strategies using MONAI transforms, and class weight balancing to address the inherent class imbalance in the dataset. The ViT achieved state-of-the-art (SOTA) performance on the MMOTU dataset with 82.98% ± 1.89% accuracy for 8-class classification and 91.15% ± 1.48% accuracy for three-class classification, establishing new SOTA results for this dataset. The low standard deviations (σ < 2%) across folds demonstrate model stability, reproducibility and reliability. The success achieved by the transformer-based architecture demonstrates the superiority of attention mechanisms over traditional convolutional approaches for ultrasound image analysis. This work contributes to the advancement of ovarian cancer point-of-care diagnostic applications for consumer healthcare devices. The proposed approach shows potential to support clinical decision-making in point-of-care settings.