KI-67 Image Analysis: A Comparative Analysis of an Organ-based Approach


Background: Ki-67 is a marker of cell proliferation that stains actively proliferating nuclei.  Some of its various indications include grading tumors, predicting prognosis, and guiding treatment in different neoplasms. Moreover, it continues to be an active area of research. Ki-67 percentage automation is a valuable tool that can standardize and speed assessment thus augment pathologists’ diagnostic workup. We present here a comparison study that analyzes the performance of two nuclear detection algorithms in neuroendocrine tumors of two organs. 

Methods: 144 glass slides from previously signed out, randomly selected pituitary adenomas (N=101) and gastrointestinal neuroendocrine tumors (GI-NETs, N=36) were scanned using Leica AT2 whole slide image scanners, and regions of interest were selected by the pathologist (Leica Biosystems, Nussloch, Germany). Automated Ki-67 percentages of those regions were generated using two nuclear detection image analysis algorithms simultaneously, a Leica base nuclear detection algorithm (A1) and an in-house nuclear detection image analysis algorithm (A2). Using the pathology reports’ manual Ki-67 percentage as the gold standard, we defined agreement as automated Ki-67 within 3% of the gold standard. Correlation coefficient and R-squared were calculated using Microsoft Excel. 

Algorithm parameters compared were the smoothing filter radius that smooths the object edges (smoothing), de-clustering of nuclei (merging), area for detectable nuclei (size), area of object to area of its perimeter ratio (compactness), object length and width ratio (elongation), weak nuclear intensity threshold (intensity threshold), and mode of intensity image threshold adjustment that allows for background noise elimination and nuclear edge detection (method), 

Compared to A1 algorithm, A2 algorithm was set to have lower values in the smoothing, merging, maximum size parameters and it was set to have higher values for compactness, elongation and intensity threshold. A1 algorithm method was set to automatically adjust the threshold based on the average pixel intensity. A2 was set to adjust the threshold based on a relatively increased upper threshold limit. 

Results: It was observed that A1 performed better in highly cellular tumors with compact cells, indistinct cell borders, lower nucleus:cytoplasmic ratio and smaller nuclei, as in pituitary adenomas. In contrast, A2 did better in tumors with distinct cell membrane, abundant cytoplasm and larger round to oval nuclei such as GI-NETs. 

For A1 algorithm, pituitary adenomas had correlation of 0.92 and R-squared of 0.84. GI-NETs had a correlation of 0.75 and R-squared of 0.56. For A2, pituitary adenomas had correlation of 0.71 and R-squared of 0.50, whereas GI-NETs had a correlation of 0.86 and R-squared of 0.74. 

Overall, discrepancy was due to cell detection and nuclear segmentation inaccuracies, tissue staining differences, annotation area under-representing overall ki-67, and deeper tissue block sections variation.  

Conclusion: Providing image analysis algorithms for computing Ki-67 percentages serves as a valuable tool that can augment pathologists’ diagnostic workup. A one-size-fits-all approach to image analysis is not as effective as taking into consideration characteristic organ-based features of tumors. Creating nuclear detection algorithms that cater to the most common distinguishing features of tumor types has the potential of improving image analysis detection rate, accuracy and performance overall.



  1. Recognize Ki-67 percentage automation as a valuable tool
  2. Appreciate the short-comings of a "one-size-fits-all" image analysis
  3. Become aware that image analysis can be further optimized when the characteristic organ-based features of tumors are taken into consideration when creating or choosing an algorithm


Presented by:


Rand Abou Shaar, MD

Informatics Fellow



Rand Abou Shaar is an enthusiastic Board certified anatomic, clinical and molecular pathologist and pathology informaticist with an inquiring mind and expertise in pathology informatics. She enjoys taking part in process improvement initiatives and volunteering.