Fig. 4

Visualization of spectral attention weights produced by the channel-wise attention module in MS-HAN for tumor tissue, aggregated across all patients. For each sample group (bag), patch-level spectral attention vectors are first aggregated within the group using normalized MIL pooling weights over tumor instances (identified by labels), yielding a group-level tumor spectral attention vector. The final curve is obtained by averaging these group-level vectors over all tumor groups. (a) Bar plot representation: Mean normalized attention weights for the 128 spectral bands spanning 397–1032 nm (color indicates magnitude; yellow: low, red: high). (b) Heatmap representation: A complementary heatmap visualization of the same aggregated weights. The resulting non-uniform band-wise distribution indicates that the model assigns systematically different importance to different spectral bands, providing quantitative and reviewer-requested interpretability of the learned spectral emphasis under weak (slice-level) supervision.