Figure S4.

Palmskin-encoded multiplex information is important for the model’s prediction accuracy.

(A) Schematic flowchart of the inputs, Vision Transformer model, and growth size classification. Four different sets of feature-subtracted palmskin images were used as inputs. (B) Representative feature-subtracted palmskin images. In “Random color #1,” SEC clonal information was preserved, but cellular texture was removed. In “Random color #2,” both clonal and cellular texture information were eliminated. In “cell-less,” cell–cell boundaries were fully removed, leaving only information related to SEC coverage. Selected cells and clones are highlighted in each image set. (B, C) Confusion matrix and F-scores of the corresponding image sets shown in (B). The confusion matrix on the far left in this panel is a duplicate of Fig 4C. Scores above 0.7 are shown in black for ease of visualization.