DeiT-MLP-mixer for the preoperative prediction of axillary lymph node involvement in breast cancer via ultrasound imaging

Abstract

Background Accurate preoperative prediction of axillary lymph node (ALN) metastasis is essential for guiding treatment decisions in breast cancer patients and avoiding unnecessary surgeries. Conventional methods, including clinical examination and imaging, often lack sufficient accuracy, whereas standard surgical procedures such as sentinel lymph node biopsy and axillary lymph node dissection (ALND) are invasive and carry potential complications. This study aims to develop a reliable, noninvasive deep learning model for predicting ALN status via grayscale ultrasound images. Methods We propose a hybrid deep learning architecture, DeiT-MLP-Mixer, which combines a data-efficient image transformer (DeiT) for global feature extraction with an MLP-Mixer-inspired classifier for final prediction. The DeiT module captures global spatial relationships, whereas the MLP-Mixer integrates local and global features through spatial and channel mixing. The model was trained and validated on a real-world dataset of 502 breast cancer patients (1,506 augmented ultrasound images), with ALN status confirmed via postoperative pathology. The performance was evaluated in terms of accuracy, precision, recall, specificity, F1 score, and loss. Compared with pretrained convolutional neural networks (CNNs) and standard transformer-based models, the proposed approach has improved accuracy and reliability, particularly in handling limited clinical datasets. Results The DeiT-MLP-Mixer achieved high performance on the test set, with an accuracy of 95.70, precision of 0.9760, recall of 0.9242, specificity of 0.9867, F1 score of 0.9494, and a loss of 0.1345. It outperformed baseline models, including Vision Transformer, DeiT, and pretrained CNNs such as InceptionV3, MobileNetV2, VGG16, ConvNext, EfficientNet, and DenseNet. Conclusions Our study presents the DeiT-MLP-Mixer as a practical and effective technique for predicting ALN status in breast cancer patients, demonstrating its effectiveness in achieving higher accuracy and specificity than conventional CNNs and standard transformer models do, even with limited clinical data. This work contributes to ongoing efforts to use artificial intelligence in the development of noninvasive tools that support clinicians in preoperative decision-making. In particular, knowing the presence or absence of ALN involvement prior to surgical planning can facilitate earlier prognostic assessment, guide the subsequent treatment course, and help reduce unnecessary invasive axillary surgical procedures.