Investigation into diagnostic agreement using automated computer-assisted histopathology pattern recognition image...

Area of Science:

• Digital pathology
• Computational pathology
• Histopathology image analysis

Background:

• The diagnostic accuracy of histopathology pattern recognition image analysis (PRIA) compared to traditional microscopic assessment remains largely unquantified.
• Evaluating PRIA's performance is crucial for its integration into clinical diagnostic workflows.

Purpose of the Study:

• To assess the agreement and reproducibility of a commercial PRIA platform against manual morphometric image segmentation and pathologist assessments.
• To identify the strengths and limitations of PRIA in analyzing diverse histopathological samples, including pulmonary metastases and teratomas.

Main Methods:

• Whole-slide images were analyzed using a commercial PRIA platform and manual morphometric segmentation.
• Statistical analyses included Passing/Bablok regression and Bland-Altman analysis to evaluate agreement and bias.
• Reproducibility was assessed using repeated measures and concordance correlation coefficients.

Main Results:

• PRIA demonstrated substantial agreement with manual segmentation for pulmonary metastatic cancer areas, with high reproducibility (CV: PRIA=7.4, manual=17.1).
• PRIA exhibited limitations with morphologically complex teratomas, leading to diagnostic inaccuracies influenced by histomorphology and algorithmic constraints.
• Method disagreement primarily impacted measurements of smaller tumor burdens (<3%).

Conclusions:

• PRIA shows promise for analyzing tissues with limited phenotypic diversity, such as pulmonary metastases, offering superior reproducibility.
• Further technical advancements and consistent pathologist input are essential to enhance PRIA's accuracy and applicability for complex histopathological diagnoses.
• PRIA's current limitations in handling diverse histomorphology necessitate careful consideration for its deployment in complex cases.