In this study, we delved into the application of deep learning models predicated on the Transformer architecture in the analysis of colon cancer diagnostic images. The experimental outcomes revealed that, for medical images featuring tissue cell samples akin to those devoid of backgrounds, the magnitude of the training dataset significantly influences the model's training efficacy. More precisely, when contrasted with the Vision Transformer model, the Swin Transformer model demonstrates the capability to attain a recognition accuracy comparable to the former on a smaller training sample set, whilst also necessitating a lesser number of training iterations. Under the condition of an ample dataset scale, subsequent to roughly 30 rounds of training, the Swin Transformer model's recognition accuracy was able to reach 100%. These insights offer valuable guidance for future decisions pertaining to the choice of suitable datasets, model frameworks, and training methodologies within the domain of small-sample-based biological tissue cell disease recognition.