Figure 8
From: Multimodal deep learning for midpalatal suture assessment in maxillary expansion
Grad-CAM visualizations of the CVM-LCR, MTM-CBCT, and MPS-CBCT modalities. Heatmaps indicate discriminative regions associated with automated decision-making, with red representing high diagnostic relevance and blue representing low relevance. The attention patterns demonstrate that the model focuses on clinically established anatomical landmarks for each modality. Grad-CAM visualization of model attention patterns. Heatmaps showing discriminative regions for each imaging modality: CVM-LCR focuses on cervical vertebrae morphology; MTM-CBCT emphasizes root development patterns; MPS-CBCT highlights posterior suture regions. Red indicates high diagnostic relevance, blue represents low relevance, confirming anatomically meaningful feature learning.
