FM-LiteLearn: A Lightweight Brain Tumor Classification Framework Integrating Image Fusion and Multi-teacher Distillation Strategies

Shengbo Tan, Ying Cai, Yang Zhao, Junjie Hu, Yuanyuan Chen, Chenxi He

Abstract: This paper presents FM-LiteLearn, an efficient and lightweight framework specifically designed for the classification of brain tumors. To address the insufficient diversity and low accuracy in image training samples, the proposed framework comprises three key modules: image fusion, model enhancement, and knowledge distillation. Firstly, we employed an image fusion technique based on Generative Adversarial Networks (GAN), specifically F-DCGAN, which integrates T1-weighted and T2-weighted brain tumor images to obtain more comprehensive tumor feature information. Additionally, we proposed an improved residual network model, T-Resnet18, which incorporates a channel attention mechanism after each residual block to enhance the recognition capability of tumor regions while reducing redundant information. Finally, a multi-teacher knowledge distillation model, MT-KD, was introduced to guide the training of T-Resnet18 using multiple large teacher models, thereby striking a balance between the number of model parameters and performance. This paper presents a novel framework for classifying Astrocytoma, Germ cell tumor, Meningioma, and Neurofibroma types. The algorithm was verified on the BT_NAGMN5 dataset, and the experimental results demonstrate a 9.4% improvement in the accuracy of brain tumor classification. Compared to traditional models and other advanced methods, FM-LiteLearn shows significant advantages in enhancing model generalization capabilities and the accuracy of brain tumor classification. The code is available at https://github.com/goblin327/FM-LiteLearn.