The flowchart of this study. It involves: (1) training a CNN-based flare prediction model using vector magnetic field data from solar active regions; (2) identifying MFRs via Grad-CAM; and (3) evaluating their physical significance through two quantitative experiments: comparing the predictive capability of parameter-based models using features extracted from MFR, SHARP, and PIL masks, and analyzing magnetic complexity via the polarity imbalance index.