Fig. 4: Visualization of attention and patch-level classification scores for LNM prediction using the UNI_CTMIL model. | npj Precision Oncology

Fig. 4: Visualization of attention and patch-level classification scores for LNM prediction using the UNI_CTMIL model.

From: Customized transformer for lymph node metastasis prediction from lung adenocarcinoma histology in a multicentric study

Fig. 4: Visualization of attention and patch-level classification scores for LNM prediction using the UNI_CTMIL model.

A, B LNM-positive cases, C, D LNM-negative cases. From left to right, each column shows the original WSI, tumor regions detected by the tumor patch detection model, the attention heatmap (final-layer transformer attention scores), and the classification heatmap (per-patch prediction probabilities from the trained model). Attention scores are normalized within each WSI. Classification scores are output probabilities ranging from 0 to 1, with higher values indicating greater likelihood of LNM.

Back to article page