Digital Ki-67 quantification in neuroendocrine tumours: comparing performance of camera-captured images and open-source platforms across imaging modalities.

Abstract

Aims: Accurate assessment of the Ki-67 proliferation index (PI) is essential for grading and prognostication of gastrointestinal well-differentiated neuroendocrine tumours (NETs). While manual counting (MC) of 500-2000 tumour cells remains the standard, digital image analysis (DIA) offers potential advantages in efficiency and reproducibility. We evaluated the comparability of open-source DIA platforms on camera-captured (CC) images and whole-slide images (WSI) for Ki-67 quantification.

Methods: Ki-67 hotspot areas of 70 NETs were photographed using a microscope-mounted camera. PI was determined by MC (gold standard) and compared with automated counts in 68 cases (two excluded owing to high background staining) using QuPath (V.0.4.4). In a randomly selected subset of 20 cases, the same hotspot areas were analysed using ImageJ, ChatGPT V.4.0 (colour-based segmentation) and the IHCexpert.com platform. Additionally, WSI files of these 20 cases were imported into QuPath for DIA; PI of identical areas were compared against static images.

Results: DIA using QuPath (on CC images) demonstrated excellent agreement with MC (intraclass correlation coefficient). Only one case showed grade reclassification (manual G1, 2.92%; DIA G2, 3.38%). In the subset analysis (n=20), comparable Ki-67 indices were observed across all digital platforms and between CC images and WSI. Grade switches from changes in Ki-67 PI were observed in two additional cases (G2 to G1 in IHCexpert.com group and G1 to G2 in ChatGPT group).

Conclusions: Our findings offer the prospect of eliminating variability in the analysis of PI estimation. Of note, CC images yield results similar to WSI, supporting broader applicability in resource-limited practice settings.