Comparing non-machine learning vs. machine learning methods for Ki67 scoring in gastrointestinal neuroendocrine
Nazanin Mola1,2, Hrafn Weishaupt3, Valentin Krasontovitsch4
1Department of Pathology, Haukeland University Hospital, Post Office Box 1400, 5021, Bergen, Norway. Nazanin.mola@ihelse.net.
Scientific Reports
|July 29, 2025
Summary
Machine learning (ML) image analysis significantly improves the accuracy of Ki67 scoring for neuroendocrine tumors compared to non-ML tools. This digital pathology approach enhances detection of tumor cells and Ki67 positivity, offering a more reliable prognostic biomarker assessment.
Area of Science:
- Digital Pathology
- Computational Biology
- Oncology
Background:
- The Ki67 score is a critical prognostic marker for neuroendocrine tumors (NETs).
- Manual Ki67 assessment is time-consuming, requiring cell counting in specific tumor regions.
- Digital image analysis offers potential for automated and efficient Ki67 scoring.
Purpose of the Study:
- To compare the performance of a machine learning (ML) tool against a non-ML tool for Ki67 scoring in NETs.
- To evaluate the accuracy of digital image analysis in detecting tumor cells and Ki67-positive cells.
- To assess the concordance of automated Ki67 scores with manual pathological assessment.
Main Methods:
- Comparison of a non-ML tool (ImageScope) and an ML tool (Aiforia Create) on Ki67-stained NET slides.
- Analysis of 10 low-grade NET cases, focusing on 8 regions per slide.
- Evaluation of cell detection (total and Ki67-positive) and Ki67 score calculation using performance metrics and interclass correlation (ICC) against manual scoring.
Main Results:
- The ML tool demonstrated superior performance (F-score 0.90) in detecting tumor cells compared to the non-ML tool (F-score 0.74).
- Higher agreement with the reference standard was observed for the ML tool in detecting tumor cells (ICC 0.91), Ki67-positive cells (ICC 0.70), and calculating the Ki67 score (ICC 0.86).
- The non-ML tool showed significantly lower agreement (ICC 0.62, 0.24, and 0.45, respectively).
Conclusions:
- Machine learning-based digital image analysis offers enhanced accuracy for Ki67 scoring in neuroendocrine tumors.
- ML tools outperform traditional non-ML methods in accurately identifying tumor cells and assessing Ki67 positivity.
- This technology streamlines prognostic biomarker assessment, improving diagnostic efficiency and reliability in neuroendocrine tumor pathology.
