Multi-modality image fusion model of AFM-derived extracellular vesicles for NSCLC subtype diagnosis

Abstract

Non-small cell lung cancer (NSCLC) comprises multiple subtypes with distinct treatment strategies, making accurate subtype classification important. Therefore, this study aims to develop a multi-modality AFM-based deep learning model to improve NSCLC subtype diagnosis because extracellular vesicles (EVs) measured by atomic force microscopy (AFM) contain nanoscale biophysical information with diagnostic potential. The proposed model introduces two key components: (i) an intra-image feature extractor (IFE), which captures parameter-specific local and global dependencies, and (ii) a cross-image feature fusion module (CFFM), which mitigates domain mismatch among heterogeneous AFM parameters. Through three fusion strategies at the image, low-level feature, and high-level feature stages, the proposed model simultaneously learns intraparameter and inter-parameter relationships. The proposed model achieved an accuracy of 0.9245 and an area under the receiver operating characteristic curve (AUROC) of 0.9804, outperforming the AFM-based conventional state-of-the-art (SOTA) model by improvements of 0.1328 in accuracy and 0.0273 in AUROC. These results demonstrate that the proposed multi-modality fusion approach achieves superior diagnostic performance for AFM-based EV analysis. Furthermore, the proposed framework has the potential to be extended to AFM-based analysis of other cell types and diseases once large-scale datasets are secured. Code is released at https://github.com/sunh124/MMIF_Model_of_AFM-derived_EV.