Fig 1.
Example results from the automatic mitosis detection algorithm.
In the images on the left side, a green cross next to an object indicates a detected mitotic figure and the numbers indicate how many observers annotated the object as a mitotic figure. The corresponding probability maps that are the output of the deep convolutional neural network are given on the right side. All local maxima with a value larger than 0.85 are considered detected mitotic figures.
Fig 2.
Modified Bland-Altman plots for the log-transformed mitotic counts of the 3 observers and the automatic method.
The black line indicates the mean difference b, and the red lines indicate the limits of agreement LA. Logarithm with base 10 was used in the transformation. The points are jittered in the horizontal direction to help better visualize overlaps.
Table 1.
Confusion matrices for the mitotic activity score for the 3 observers and the automatic method.
Table 2.
Dice similarity coefficients (D) for all pairs of observers and the automatic method for the entire dataset.
Fig 3.
Distribution of the Dice similarity coefficients for individual cases and for all pairs of observations.
The blue line indicates the median. The analysis is limited to the cases for which ground truth mitotic figures are available from all three observers (N = 84).
Fig 4.
Top (first eight rows): Objects that were annotated as mitotic figures by all three observers.
Bottom (last eight rows): Objects that were annotated as mitotic figures by only one observer. The numbers in the top-right corner of the patches indicate the observer that annotated the object. A green bar at the bottom of the image patch indicates that the object was detected by the automatic method.
Fig 5.
Size of the mitotic figures with poor and good agreement.
The mitotic figures annotated by only one observer were on average smaller than those for which all three observers agreed.