(a-b) Pairwise classification of nuclei from breast cell lines in the latent space.
(a) Classification results in the latent space using a linear model. Top: training and test loss curves for each pairwise comparison (HME1, MCF10A, MCF7, MDA-MB231). Middle: training and test accuracy curves for each pairwise comparison. Bottom: Table of best training (red) and test (green) accuracy for each classification task. (b) Same as (a) but using a two-layer feedforward neural network. For all tasks, the sizes of the training and validation datasets were 550 and 60 respectively. (c-e) Pairwise classification of nuclei from breast cell lines using VoxNet. (c) Accuracy and (d) loss curves for each pairwise comparison (HME1, MCF10A, MCF7, MDA-MB231). (e) Table of best training (red) and test (green) accuracy for each classification task. (f-k) 4-way classification results for nuclei from breast cell lines. Training and test loss curves for 4-way classification task of HME-1, MCF10A, MCF7 and MDA-MB231 nuclei using a linear model (f) and 2-layer feedforward neural network (g). (h-i) Confusion matrices for the classification tasks in (f-g). Each entry (X/Y) in row “A” and column “B” indicates that X nuclei of class “A” were classified as “B” in the training set and Y nuclei of class “A” were classified as “B” in the test set. (j-k) Same as (g,i) but for the 4-way classification task in the original image space using a deep convolutional neural network.
(TIF)