In recent years, there has been a significant increase in the number of observed astronomical spectra. This has led to greater demands for automatic classification and analysis of spectra in large-scale spectroscopic surveys. In this study, we introduce a multi-scale feature fusion-based stellar spectral classification model (MSFnet) that leverages the strengths of Convolutional Neural Networks (CNNs) in classification tasks and employs a multi-scale feature fusion module to extract spectral features at various scales for the prediction of stellar spectral types. The proposed MSFnet architecture consists primarily of a multi-scale feature fusion module and a CNN with four convolutional layers, two max-pooling layers, and one fully connected layer. To mitigate overfitting, dropout is incorporated into the model, enhancing its robustness by reducing dependence on specific local features. The dataset employed in this study is derived from the LAMOST DR9 database. Before training, data preprocessing is performed, which includes uniform resampling of spectra and min-max normalization. The experiment uses Python 3.9 and the PyTorch deep learning framework to build the network. The experimental section is divided into two parts: the first part investigates the relationship between the number of layers in the CNN, the number of feature maps, and the classification accuracy; the second part compares the performance of the proposed MSFnet model and the Resnet18 model using evaluation metrics such as precision (P), recall (R), and F1 score. Both models training, validation, and test sets are split according to a 6∶2∶2 ratio to maintain consistency in training samples. Results demonstrate that the highest accuracy is achieved using a CNN with four convolutional layers and 16 feature maps. Based on this finding, we propose the MSFnet model, which combines the feature fusion module with the CNN. Compared to the 18-layer residual neural network model, the MSFnet model has a more straight forward structure and performs similarly on the evaluation metrics. The performance in the metrics above is on par with that of the Resnet18 model. Furthermore, it demonstrates superior classification efficacy for spectra types A, F, and K, accompanied by enhanced speed. With an accuracy of nearly 97% on the test set, the MSFnet model outperforms traditional CNN and Resnet18 models, indicating its potential to improve the accuracy of automatic spectral classification significantly.