The paper presents TSJNet, a novel network for multimodal MRI data fusion that combines information from different sensors to improve spinal tumor diagnosis. The architecture includes a fusion module with detection and segmentation subnets, while an LSFE module with a two-branch design improves the interaction between the functions of different modalities. The model was tested on four public datasets (MSRS, M3FD, RoadScene, LLVIP) where it achieved an average improvement of +2.84% in object detection (mAP@0.5) and +7.47% in semantic segmentation. Comparison with classic ML methods (DWT + SVM, LBP + SVM) and modern deep networks confirmed its superiority. On the MSRS set, 5-fold cross-validation showed a performance of 78.21 ± 1.02 mAP and 71.45 ± 1.18 mIoU. Model complexity analysis demonstrated efficiency in parameters, FLOPs, and inference time. TSJNet effectively combines task supervision and modality interaction for high-quality fused outputs suitable for real-world applications.