As deep learning models become essential tools in medical image analysis for lesion detection and disease diagnosis, existing course designs often fail to help students bridge the gap between theoretical knowledge of complex architectures, specifically convolutional neural networks (CNNs) and vision Transformers (ViTs), and their practical application in clinical scenarios. To address this, a comprehensive experimental design is proposed for classifying vocal fold leukoplakia using narrow-band imaging (NBI) laryngoscopy images. This experiment guides students in constructing a hybrid deep learning model, LeukoplakiaNet, which integrates the local feature extraction capability of CNNs with the global dependency modeling ability of ViTs. Through this experiment, students transition from theoretical learning to solving practical engineering problems, specifically the accurate classification of precancerous laryngeal lesions.

The experiment was implemented using Python and the PyTorch deep learning framework, following a complete pipeline of five core modules: data preprocessing, data augmentation, network pre-training, hybrid model training, and model testing. A clinical dataset of 665 NBI laryngoscopy images was obtained from the Second Affiliated Hospital of Xi’an Jiaotong University, covering six pathological categories: normal tissue, inflammatory keratosis, mild dysplasia, moderate dysplasia, severe dysplasia, and squamous cell carcinoma. All images were resized to 224×224 pixels and split into training and testing sets at a 4∶1 ratio. The core innovation is LeukoplakiaNet, a hybrid model that uses ResNet50 as the backbone for extracting local pathological features. To overcome the limited global receptive field of CNNs, a ViT was employed as a global attention network. A novel spatial-channel attention (SCA) module was designed to bridge the CNN and ViT components, incorporating a spatial attention submodule to highlight key lesion areas and a channel attention submodule to identify critical feature channels. This collaborative design enables effective fusion of local details with global semantic context. To address the small sample sizes common in medical imaging, transfer learning was incorporated by initializing the model with weights pretrained on ImageNet-1K. Data augmentation techniques (e.g., flipping, rotation, contrast adjustment) were applied to the training set to prevent overfitting.

In the six-category classification task for vocal fold leukoplakia, LeukoplakiaNet achieved an overall accuracy of 93.79%, representing a marked improvement over single-architecture models: it outperformed a standalone ResNet50 by 12.40% and a standalone ViT by 5.42%. The model achieved 100% recognition accuracy for both mild and severe dysplasia categories. Sensitivity and specificity for all six categories remained above 0.87. The experiment also incorporated Grad-CAM visualization, enabling students to generate heatmaps of the model’s decision-making process. These heatmaps confirm that the model focuses on clinically relevant areas, such as leukoplakia regions and abnormal microvessels, consistent with professional medical diagnoses.

This experiment constructs an integrated teaching workflow of “data preprocessing–model training–performance evaluation–result visualization.” Through this workflow, students deepen their understanding of hybrid deep learning models and master techniques for medical image processing. The design cultivates scientific research aptitude, enhances practical engineering abilities, and provides a solid foundation for future research in intelligent medical diagnosis.