Paroxysmal atrial fibrillation (PAF) is difficult to detect because patients often remain in sinus rhythm between arrhythmic episodes. This study proposes a Multi-scale Temporal-Graph Fusion Network (MTGF-Net) for PAF detection using high-density 96-lead electrocardiograms (ECGs). The framework combines multi-scale temporal feature extraction with graph-based modeling of inter-lead spatial relationships. A 96-lead ECG dataset was constructed from 65 participants, including 31 patients with PAF and 34 healthy controls. A total of 5,561 valid ECG segments were retained. Compared with the Multi-CNN and ResNet baselines, the Graph Convolutional Network (GCN)-based variant, MTGF-GCN, achieved the best overall performance, with an accuracy of 85.26%, an AUROC of 0.931, and a sensitivity of 81.79%. In addition, an approximate 12-lead configuration achieved lower performance than the full 96-lead input. These findings support the potential of high-density ECG and spatial modeling for identifying PAF-related patterns during sinus rhythm.