Model: Random forest

Accuracy

Random forest accuracy on validation data was 0.754, and on test data 0.75.

Accuracy per category

Across classes, the mean and median precision were 0.7 and 0.71 respectively, recall 0.7 and 0.76, and accuracy 0.96 and 0.96.

Confusion matrix

Each element \(M_{ij}\) in the confusion matrix \(M\) shows the number of observations from class \(i\) which were predicted as class \(j\). The confusion matrix in the figure below has been normalized by class support size (each row has been divided by the number of elements in that class).

To see misclassification patterns more clearly, the diagonal of the confusion matrix has been excluded in the following figure.

Protein classification accuracy

Across proteins, the mean and median accuracy were 0.954 and 1.000, respectively.

Protein classification accuracy (at least 10 cell images)

Across proteins with at least 10 cell images, the mean and median accuracy were 0.964 and 1.000, respectively.

Misclassifications - false negatives

Original and scaled cell images, respectively.

## [1] "cell periphery, false negative rate: 0.262"

## [1] "cytoplasm, false negative rate: 0.154"

## [1] "endosome, false negative rate: 0.527"

## [1] "er, false negative rate: 0.251"

## [1] "golgi, false negative rate: 0.238"

## [1] "mitochondrion, false negative rate: 0.222"

## [1] "nuclear periphery, false negative rate: 0.172"

## [1] "nucleolus, false negative rate: 0.156"

## [1] "nucleus, false negative rate: 0.167"

## [1] "peroxisome, false negative rate: 0.628"

## [1] "spindle pole, false negative rate: 0.438"

## [1] "vacuole, false negative rate: 0.4"

Misclassifications - false positives

## [1] "cell periphery"

## [1] "cytoplasm"

## [1] "endosome"

## [1] "er"

## [1] "golgi"

## [1] "mitochondrion"

## [1] "nuclear periphery"

## [1] "nucleolus"

## [1] "nucleus"

## [1] "peroxisome"

## [1] "spindle pole"

## [1] "vacuole"

Common mistakes (top 5)

## [1] "True class: peroxisome (total images 164), predicted as: spindle pole, count: 82 (50.0% error)"

## [1] "True class: endosome (total images 689), predicted as: mitochondrion, count: 156 (22.6% error)"

## [1] "True class: spindle pole (total images 781), predicted as: endosome, count: 168 (21.5% error)"

## [1] "True class: vacuole (total images 587), predicted as: nuclear periphery, count: 105 (17.9% error)"

## [1] "True class: vacuole (total images 587), predicted as: mitochondrion, count: 69 (11.8% error)"