Multi-Class Brain Lesion Classification Using Deep Transfer Learning With MobileNetV3

Abstract

Diagnosing brain tumors is challenging for radiologists because of the significant similarities between the tumor types. Deep learning models lack sufficient data to effectively learn the patterns of different tumors, leading adopting of transfer learning as a successful approach. However, many existing models used for this purpose are complex and involve numerous parameters and layers. In this study, we employed a lightweight MobileNetV3 model to extract features, specifically designed for mobile CPU usage, to transfer knowledge. We then design our model for brain lesion classification by incorporating lightweight DepthWise and PointWise blocks. A combination of three datasets with identical image structures is utilized, and compared its classification performance with both pre-trained and fine-tuned methods. The proposed model achieves an accuracy of 91%, outperforming other pre-trained and fine-tuned methods. Furthermore, we conduct separate accuracy assessments for each dataset, demonstrating superior performance compared to existing methods. Specifically, our model achieves an accuracy of 91% on the NINS 2022 dataset and 94% on the SBE-SMU dataset.