Figure 5.

Figure 5. Grad-CAM recognizes key cellular features important for the model’s decisions.

(A) Schematic flowchart of the generation of CAM images for quantification. (B, C) Representative palmskin images and corresponding CAM images. Two individuals from the small-sized and large-sized categories are shown as indicated. (D) Violin plot of all the largest hotspots from each sub-cropped CAM image captured from small- and large-sized fish (red line, median; blue line, quartiles; dotted black line, mean). Percent differences are shown above the horizontal lines for intergroup comparisons. n = number of hotspots examined. (E) Schematic flowchart of the inputs, Vision Transformer model, and growth size classification. Of note, instead of palmskin images, bright-field images and trunk surface area were used as inputs. The input bright-field images were rescaled to a uniform size of 1,288 × 214 pixels to remove size-related information. (F) Confusion matrix and F-scores of the training results. Scores above 0.7 are shown in black for ease of visualization. (G) Representative CAM images overlaid on the bright-field images. Six individuals from the small-sized category are shown.