Table 1.
Summary of symbols, meanings, and tensor shapes used in the proposed method.
Fig 1.
Schematic of the MedSpectralNet architecture with parallelizable convolutional operations.
Fig 2.
Structure of the SpectralFlow Module.
Fig 3.
ContextGate block with three pathways: spatial (depthwise conv + SpectralFlow), gating (pointwise conv + GELU), and identity (pointwise conv).
Fig 4.
Training and test loss (left) and accuracy (right) curves for MedSpectralNet on OrganCMNIST dataset over 200 epochs.
Fig 5.
Confusion matrix for MedSpectralNet on OrganCMNIST dataset showing class-wise prediction distribution across 11 organ categories.
Fig 6.
Multi-class ROC results of MedSpectralNet on the OrganCMNIST dataset.
Fig 7.
Multi-class PRC results of MedSpectralNet on the OrganCMNIST dataset.
Fig 8.
Training and test loss (left) and accuracy (right) curves for MedSpectralNet on OrganSMNIST dataset over 200 epochs.
Fig 9.
Confusion matrix for MedSpectralNet on OrganSMNIST dataset showing class-wise prediction distribution across 11 organ categories.
Fig 10.
Multi-class ROC results of MedSpectralNet on the OrganSMNIST dataset.
Fig 11.
Multi-class PRC results of MedSpectralNet on the OrganSMNIST dataset.
Fig 12.
Grad-CAM visualization for ContextGate Block on BreastMNIST.
Class 0 (normal) shows distributed attention suggesting holistic assessment, while Class 1 (abnormal) focuses on suspicious regions, reflecting adaptive gating behavior.
Fig 13.
Grad-CAM for SpectralFlow Module on BreastMNIST.
Visualizations show precise, localized activations on diagnostically relevant regions in Class 1, while Class 0 displays more distributed patterns.
Table 2.
Ablation study of different optimizers and augmentation strategies on the BreastMNIST dataset. The best performance is highlighted in bold.
Table 3.
Accuracy comparison of MedSpectralNet with benchmark models across selected datasets.
Table 4.
Computational efficiency comparison on medical image classification. FLOPs measured in giga floating-point operations (G).
Fig 14.
Parameter size required by MedSpectralNet in comparison to benchmark models.