Biomedical image analysis, such as tissue and disease classification by using small-scale data, is a challenging and meaningful problem for clinical diagnosis. In previous works, most classical methods are using 2D convolution to extract feature from a single direction without making information of z-axis into consideration and haven’t considered that information representations are different in every view direction. In this letter, an attention mechanism-based model is proposed to extract and utilize feature more efficiently under the consideration of various information representation from direction to direction. To reach this goal, the work mainly consists of three parts of characteristics. The first characteristic is a plug-and-play structure design with parallel feature extractor in the different axial directions of which is dedicated to solving the problem of losing spatial structure feature by traditional 2D convolution. The second characteristic is intensive skip connection blocks employed to optimize the phenomenon of insufficient transit of information flow which cause the deep layer to fail in learning biomedical attribute. The last step is a designed spatial attention module, which’s applied to capture the long-range dependencies , suppress irrelated information and enhance significant spatial feature globally. The evaluation on various dataset shows that proposed model can achieve a competitive performance with comparison to other existed methods and reach the best 85.18 accuracy on average.