Results
We find that our best models for ResNet50, Xception, and YOLOv5 all have high overall accuracies and perform well in the classes with the most images. All models started to perform worse on the classes with lower image counts, which was to be expected. Below you can find bar charts visualizing models’ performances on the classes with the most images.
From the bar charts, we see that YOLOv5 and Xception have an accuracy of 0.96 for myeloblasts. Myeloblasts are one of the most relevant cells in an AML Diagnosis (DiNardo et al., 2016).
In comparing the results, there are a few notable findings. One is that for the myeloblast class, Xception and YOLOv5 have an equally high precision of .96 (compared to ResNet50’s precision of .93) but YOLOv5 has a lower recall (.86) than Xception (.95). Recall is calculated as True Positives divided by the sum of True Positives and False Negatives. Low recall means a lot of false negatives, which means a lot of myeloblasts are not being recognized as myeloblasts by YOLOv5. Because this is an important class for AML diagnosis, and because Xception performs either the best or equally as good as the other models for every large class, Xception is our best overall model.
From the bar charts, we see that YOLOv5 and Xception have an accuracy of 0.96 for myeloblasts. Myeloblasts are one of the most relevant cells in an AML Diagnosis (DiNardo et al., 2016).
In comparing the results, there are a few notable findings. One is that for the myeloblast class, Xception and YOLOv5 have an equally high precision of .96 (compared to ResNet50’s precision of .93) but YOLOv5 has a lower recall (.86) than Xception (.95). Recall is calculated as True Positives divided by the sum of True Positives and False Negatives. Low recall means a lot of false negatives, which means a lot of myeloblasts are not being recognized as myeloblasts by YOLOv5. Because this is an important class for AML diagnosis, and because Xception performs either the best or equally as good as the other models for every large class, Xception is our best overall model.