This study presents a rapid identification method for isocitrate dehydrogenase (IDH) mutation status in glioma by integrating terahertz time-domain spectroscopy (THz-TDS) with deep learning. Absorption coefficient data of glioma tissues were acquired in the 0.2–1.4 THz frequency range, followed by Savitzky–Golay smoothing and Z-score normalization to construct an 82-dimensional feature dataset. To improve classification performance, the synthetic minority oversampling technique (SMOTE) was employed to augment the training data. The spectral features were then transformed into four types of two-dimensional representations: Gramian Angular Summation Field (GASF), Gramian Angular Difference Field (GADF), Markov Transition Field (MTF), and Recurrence Plot (RP). Three convolutional neural network (CNN) architectures were developed for comparative analysis: single-input CNN, front-end fusion CNN (FFCNN), and mid-level fusion CNN (MFCNN). The results demonstrated that MFCNN achieved the highest performance when fusing GASF and GADF images, yielding an AUC of 0.907, outperforming all other configurations. However, the inclusion of MTF images led to a performance decline, potentially due to feature conflicts introduced by modality differences.