In the immunotherapy process, a machine vision algorithm exhibits an efficient next-generation model for navigating the complex tumor microenvironment with CAR-T cells. With the integration of image-based analysis into the real-time processing algorithm, the system is able to compute spatial guidance for the immune cell, enabling it to detect, eliminate, and infiltrate cells. The variation between computational vision and cellular therapy needs to overcome the issues in the physical and biological barriers of tumors. Hence, in this paper, an effective Fejer Kernel Entropy Masked R-Convolutional Neural Network (FEM-R-CNN) was constructed. The proposed FEM-R-CNN model performs pre-processing of the CAR-T cell using the Fejer Kernel, and segmentation is performed with the entropy model. With the estimated segmentation, the Single Shot Detector (SSD) is employed for the CAR-T cell, and classification is computed using the masked R-CNN for the immunotherapy. Experimental results demonstrate that FEM-R-CNN achieves a cancer cell detection accuracy of 96.2%, a segmentation Intersection over Union (IoU) of 0.86, and a classification accuracy of 94.8%, outperforming traditional models such as U-Net and standard Mask R-CNN by over 5% across key metrics. The model improves signal-to-noise ratio by 46.4% and reduces false positive rates by 53.3%, enabling more precise CAR-T cell navigation. Immune response analysis revealed a CAR-T cell density of up to 150 cells/mm², with a 50% proliferation rate and 72% tumor cell apoptosis, indicating effective immune activity monitoring. The inference time remains competitive at approximately 70 ms per image, supporting near real-time applications.