Accurate tumor cell detection is essential for clinical diagnosis and treatment. Traditional methods rely on manual analysis of fluorescence microscopic images, which are inefficient, costly, and difficult to scale for large-scale applications. To address these limitations, we propose FMIC-AI, a novel framework based on self-supervised anomaly detection (SSAD). FMIC-AI integrates general microscopy techniques, an SSAD-based network, the CellPose segmentation model, and class activation maps for precise localization. It eliminates manual annotations by leveraging only normal cell images for training. The framework employs an improved Vision Transformer (ViT) trained with self-supervised contrastive learning, enabling robust detection of abnormalities. Additionally, CellPose is fine-tuned in a self-supervised manner to reduce dependency on labeled data, while Grad-CAM++ is combined with segmentation masks to achieve accurate tumor cell localization. Experimental results demonstrate that FMIC-AI achieves an AUC of 0.934 and a recall rate of 0.89 on the test set, significantly outperforming traditional methods by up to 15% in key metrics. This study provides an efficient and scalable solution for tumor cell detection in fluorescence microscopic images, offering promising applications in clinical pathology.