FIG. 2.

A 3D FLI-LeNet CNN model for classifying glycolytic from oxidative MCF7 cells from NAD(P)H TPSF images. (a) The structure of the FLI-LeNet CNN model for predicting cancer cells using glycolysis and cancer cells using OXPHOS based on the NAD(P)H TPSF images. (b) Validation loss and training loss by epoch for FLI-LeNet models trained with different datasets (Org: original TPSF images, MD: down-sampled TPSF images with the mean filter, MEDD: down-sampled TPSF images with the median filter). Solid lines represent validation loss, and dashed lines represent training loss. (c) Validation accuracy and training accuracy by epoch for FLI-LeNet models trained with different datasets. Solid lines represent validation accuracy, and dashed lines represent training accuracy. (d) t-SNE visualization obtained from the last activation map of the FLI-LeNet model of the test data of the model trained with the original NAD(P)H TPSF images. Each dot corresponds to one cell based on its representation in the last activation layer of the pre-trained FLI-LeNet after fine-tuning. Red data points represent cells using OXPHOS, and green data points represent the cells using glycolysis. (e) Representative ROC curves of FLI-LeNet models trained with original NAD(P)H TPSF data (Org) and down-sampled data (MD, MEDD) for predicting glycolysis or OXPHOS of MCF7 cells within the test datasets.