Enhancing CAR-T Cell Tumor Targeting via Advanced Computational Perception Networks for Improved Recognition in Heterogeneous Tumors

In cancer treatments, the efficacy of Chimeric Antigen Receptor T (CAR-T) cell therapy is affected in heterogeneous tumors due to ambiguous tumor boundaries, morphological variability, and similarity between tumor and non-tumor tissues in medical imaging. Accurate tumor localization and classification are crucial for optimizing CAR-T targeting and therapeutic success. Traditional segmentation networks struggle with intensity similarity, shape variability, and contextual complexity in heterogeneous tumors. Further, robust classification of tumor regions using limited medical data remains a key challenge. We propose a dual-component Computational Perception Architecture composed of a novel segmentation-classification framework. The segmentation backbone is a U-Net enhanced with a Visual Perception Module (VPM) for ROI-level feature refinement. Multi-Head Self-Dilated Attention (MHSDA) in the encoder to capture multi-scale dependencies. ResNet50 with Dense Attention Modules in skip connections for improved feature continuity. Group Receptive Large Kernel (GRLK) Blocks for diverse receptive field decoding. The classification network utilizes edge-perception, morphological, and positional images, and segmentation maps. Deep ensemble learning for decision robustness and transfer learning to boost generalization on breast cancer labeled datasets. The proposed method is tested on the publicly available PBC and CAR-T datasets from Kaggle. The research model achieved a Dice Score of 0.901, an IoU of 0.856, a Precision of 0.882, a Classification Accuracy of 93.7%, and an F1-Score of 0.915. These outcomes show the superior capacity for precision tumor detection and classification, thus offering a potent computational aid in enhancing the targeting precision of CAR-T therapies.