Cellular micronucleus detection plays an important role in pathological toxicology detection and early cancer diagnosis. To address the challenges of tiny targets, high inter-class similarity, limited sample data and class imbalance in the field of cellular micronucleus image detection, this paper proposes a lightweight network called MobileViT-MN (Micronucleus), which integrates a multilayer perceptual attention mechanism. Considering that limited data and class imbalance may lead to overfitting of the model, we employ data augmentation to mitigate this problem. Additionally, based on domain adaptation, we innovatively introduce transfer learning. Furthermore, a novel Deep Separation-Decentralization module is designed to implement the reconstruction of the network, which employs attention mechanisms and an alternative strategy of deep separable convolution. Numerous ablation experiments are performed to validate the effectiveness of our method. The experimental results show that MobileViT-MN obtains outstanding performance on the augmented cellular micronucleus dataset. Avg_Acc reaches 0.933, F1 scores 0.971, and ROC scores 0.965. Compared with other classical algorithms, MobileViT-MN is more superior in classification performance.