Emerging evidence points to the cerebellum as a key player in working memory and executive function, yet its structural role within the fronto-thalamo-cerebellar (FTC) circuitry remains underexamined. Leveraging the expansive ABCD dataset, this study evaluated the predictive capacity of FTC anatomy—specifically gray matter volume (GMV) and fractional anisotropy (FA)—in modeling working memory (2-back accuracy) and attention (0-back accuracy). Classical models such as Bayesian regression, SVR, and feedforward neural networks achieved a peak R² of 0.062, with FTC-based predictions significantly outperforming fronto-parietal counterparts (p < 0.05). Building on this, we developed a multi-stage deep learning framework incorporating 3D CNNs, self-attention blocks, and squeeze-and-excitation mechanisms. These models pushed predictive R² to 0.079 while offering interpretable attention maps that consistently highlighted cerebellar subregions. Together, these results underscore the FTC circuit’s robust involvement in working memory and demonstrate how deep learning can reveal spatially grounded, anatomically meaningful patterns from structural MRI.