Token Mixing for Breast Cancer Diagnosis: Pretrained MLP-Mixer Models on Mammograms

Abstract

Breast cancer remains a leading cause of mortality among women, necessitating accurate and computationally efficient diagnostic solutions. Deep learning, particularly convolutional neural networks (CNNs), has significantly advanced mammographic analysis by automating feature extraction and improving early detection. However, CNNs rely on localised feature extraction, limiting their ability to capture long-range dependencies essential for robust classification. This study introduces and evaluates the effectiveness of pre-trained MLP-Mixer models using transfer learning as an alternative to CNN-based approaches, utilising their token-mixing and channel-mixing mechanisms to integrate local and global spatial features in mammograms. Four MLP-Mixer variants (B/16, L/16, B/32, and L/32) were systematically assessed on three benchmark datasets: CBIS-DDSM, INbreast, and MIAS. The results demonstrate that MLP-Mixer models, particularly those with smaller patch sizes (L/16 and B/16), consistently achieve state-of-the-art accuracy and sensitivity, while also offering 30 – 50% faster inference times compared to leading CNNs such as ResNet and DenseNet. These models demonstrate strong generalisation across multiple benchmark datasets and strike an effective balance between diagnostic accuracy and computational efficiency, which are essential requirements for clinical deployment. Their performance underscores the importance of fine-grained feature extraction in mammographic analysis. Comparative results indicate that MLP-Mixer models offer a compelling alternative to conventional CNNs by efficiently capturing both local and global dependencies without the high computational demands of deep convolutional network architectures. These findings highlight the promise of token-based models for AI-assisted breast cancer diagnosis and suggest that MLP-Mixer architectures are well-suited for real-time medical imaging applications. By enabling direct global spatial interaction, reducing architectural complexity, and improving diagnostic precision across varied imaging conditions, MLP-Mixers offer a computationally efficient alternative to traditional CNNs without compromising accuracy.