Abstract: Accurate diagnosis and subtyping of non-small cell lung cancer (NSCLC) are crucial for providing patient-specific precision treatment. However, the inherent tumor heterogeneity of NSCLC leads to significant morphological variations within the same subtype and similarities across different subtypes, presenting substantial challenges for pathologists. To address this issue, this study proposes a novel computer-aided diagnostic framework that integrates multi-scale feature extraction and fusion through multi-instance deep learning. The proposed method aims to effectively leverage the heterogeneous information presented in pathological whole-slide images (WSIs) to improve the accuracy of NSCLC subtype classification. Initially, the framework performs multi-scale sampling and feature extraction from WSIs at various levels, such as cellular and tissue levels, to capture both local and global contextual information. Subsequently, a vision transformer network is employed to model the complex dependencies among instances of varying granularity, facilitating end-to-end fusion of the extracted features for accurate classification. Furthermore, we introduce an attention-based instance loss function that adaptively weighs the contribution of each instance based on its discriminative power, providing additional supervision to enhance the classification performance of the model. We evaluat our method on a large public dataset containing 1 674 H&E-stained pathological slide images of NSCLC. The experimental results demonstrate that our multi-scale fusion method effectively leverages the rich information in multi-grained pathological data, significantly outperforming single-scale approaches in NSCLC subtype classification accuracy. Moreover, the method's attention heatmaps offer interpretability and allow for intuitive assessment of individual sample classification quality, serving as a quantitative analytical tool for further model refinement and validation. In conclusion, the proposed multi-scale multi-instance learning framework provides a powerful and interpretable solution for accurate NSCLC subtype classification, which has the potential to assist pathologists in making more reliable diagnostic decisions and ultimately improve patient care.