Figure 2
From: Automated classification of estrous stage in rodents using deep learning
EstrousNet accuracy is comparable to human experts. (A) Validation accuracy curves for EstrousNet trained using four different base architectures: ResNet-50, Inception v3, VGG-19, and MobileNet v2. All networks were trained on EstrousBank images. Mean validation accuracy across 3 training epochs (E1, E2, and E3). (B) Schematic of the EstrousBank split for training, validation, and test sets. By percentage, this split is 80%, 10%, and 10%, respectively. (C) Breakdown of EstrousBank by stain and stage. Stains from left to right are hematoxylin and eosin (HE), Shorr stain (SH), Giemsa stain (GE), crystal violet (CryV), and cresyl violet (CreV). The complete bank consists of n = 12,719 cytology images. (D) Confusion matrix of EstrousNet classifications, represented here as a heatmap, with consensus from benchmark classification acting as our ground truth. Numbers represent the number of images classified for each stage, from a comparison set made up of 400 images [100 images from diestrus (D), proestrus (P), estrus (E), and metestrus (M)]. Confusion matrix of human classification, represented as a heatmap, with ground truth stages as described previously. (F) Average test accuracy distributions in each estrous stage for EstrousNet vs. human classifications. EstrousNet distributions are identified by a continuous line while human classifications are identified by a dotted line. Distributions were created by bootstrapping data over 5000 iterations, sampling with replacement. Error bars are 25th (75th) percentiles minus (plus) the interquartile range (75th percentile minus 25th percentile). Asterisks indicate significance as determined by Fisher’s Exact Test; diestrus: odds ratio = 0.68, 95% confidence interval = 0.55–0.83, p = 1.2 × 10–5, proestrus: odds ratio = 0.68, 95% confidence interval = 0.55–0.83, p = 0.075, estrus: p = odds ratio = 0.68, 95% confidence interval = 0.55–0.83, p = 0.84, metestrus: odds ratio = 0.68, 95% confidence interval = 0.55–0.83, p = 0.60. Across all stages accuracy was significantly different, with odds ratio = 0.68, 95% confidence interval = 0.55–0.83, p = 2.1 × 10–4, Fisher’s Exact Test. (G) Venn diagram of the overlap between human expert coders, with a total of 400 classifications for each coder.
