Fully Automated Brain Tumor Volume Estimation via Multiparametric MRI and Constrained Deep Convolutional Networks

Prompt and precise volumetric assessment of malignant brain tumors via magnetic resonance imaging (MRI) remains a cornerstone for effective neurosurgical planning. This study introduces a computationally streamlined, cost-effective paradigm leveraging a tailored deep convolutional neural network topology optimized for semantic tumor segmentation and physical volume quantification. Utilizing the multi-institution BraTS cohort encompassing exactly 369 patient scans, we synthesized three imaging sequences—contrast-enhanced T1-weighted (T1ce), T2, and FLAIR—into a unified multi-parametric three-channel tensor representation. To ensure clinical accessibility, the network was optimized across 15 epochs (825 iterations) to operate exclusively under a central processing unit (CPU) architecture (requiring 236 minutes of training time and zero GPU dependency), bypassing resource-heavy hardware infrastructures. Comparative validation demonstrates that standard single-channel slice processing exhibits catastrophic localization failure, whereas the proposed multi-parametric framework achieves a robust validation accuracy of 89.72%. Furthermore, by integrating a dedicated Equation of Calibration and Volumetric Integration (ECIV) within a proprietary graphical application, "Final Tumor Pro v3," pixel-level classification probabilities are directly mapped to physical volume (mm3). The model’s performance scales progressively from a stable Dice score of 68.39% within highly ambiguous, low-contrast boundary topologies to a peak efficiency of 95.60% across core tumor slices, validating its profound resilience and clinical utility under stringent hardware constraints.