KD_ConvNeXt: knowledge distillation-based image classification of lung tumor surgical specimen sections

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Lung cancer is a leading cause of cancer-related deaths globally.
• Accurate subtyping of lung tumors is critical for effective diagnosis and treatment planning.
• Histopathological analysis of lung tumor images is essential for identifying specific subtypes.

Purpose of the Study:

• To develop an AI-driven method for classifying specific subtypes of lung tumors using histopathological images.
• To improve the accuracy and efficiency of lung cancer subtyping in clinical practice.
• To construct a clinical dataset of lung tumor histopathological images for research.

Main Methods:

• A teacher-student network architecture utilizing knowledge distillation (KD_ConvNeXt) was proposed.
• The student network (ConvNeXt) learned from the intermediate features of a teacher network (Swin Transformer).
• This approach enhanced feature extraction and model fitting, with the teacher providing soft labels to address class imbalance.

Main Results:

• The KD_ConvNeXt model achieved a classification accuracy of 85.64% on a clinical lung tumor image dataset.
• An F1-score of 0.7717 was obtained, outperforming other advanced image classification methods.
• Experiments demonstrated the effectiveness of the proposed knowledge distillation approach.

Conclusions:

• The KD_ConvNeXt model shows significant promise for accurate lung tumor subtyping from histopathological images.
• This AI-based approach can assist clinicians in diagnosing and treating lung cancer.
• Knowledge distillation effectively improves the performance of deep learning models in medical image classification.