Head-to-Head Evaluation of the Cancer Ontology Supervised Multimodal Orchestration (COSMO) AI System Versus Pathologist-Only Review

Overview

This study evaluates the diagnostic performance of the Cancer Ontology Supervised Multimodal Orchestration (COSMO) AI system for cancer subtype classification and compares it head-to-head with pathologist-only review. Pathologists will independently review de-identified whole-slide images derived from up to 300 patients across three anatomical sites (brain, lung, kidney) and provide diagnostic assessments. In parallel, COSMO will process the same cases offline to generate independent predictions, enabling direct comparison of diagnostic accuracy between human experts and the AI system. The study will characterize the diagnostic accuracy of COSMO and pathologists, inter-observer agreement, and variations in performance across anatomical sites and cancer types with different incidence rates. Results will establish how COSMO compares to pathologists on identical cases and will inform the development of AI-assisted diagnostic systems in clinical practice.

Study Rationale and Background Diagnostic accuracy in cancer subtype classification varies significantly among pathologists due to differences in expertise, experience, and access to diagnostic resources. The emergence of AI systems in pathology offers the potential to enhance diagnostic performance and consistency in cancer classification. However, direct empirical comparisons of AI-based predictions and pathologists' diagnostic performance on identical cases remain limited in the literature. Study Aims This head-to-head comparative study aims to: (1) evaluate the diagnostic performance of the COSMO AI system in cancer subtype classification across multiple anatomical sites; (2) characterize the diagnostic accuracy of experienced pathologists on the same cases; (3) directly compare diagnostic performance metrics between COSMO and pathologists; and (4) examine concordance patterns and performance variation by anatomical site, cancer incidence category, pathologist experience, and case complexity. Study Setting and Participants The study will involve up to 25 board-certified pathologists with 3 to 10+ years of diagnostic experience, recruited from institutions across North America, Europe, and the Asia-Pacific region. Participating pathologists will have domain expertise in neuropathology, pulmonary pathology, urologic pathology, or general anatomical pathology. Cases and Stratification The study will employ de-identified archival whole-slide images representing up to 300 patients with confirmed reference diagnoses, including 100 brain cancers, 100 lung cancers, and 100 kidney cancers. Cases will be stratified by cancer type and incidence category (common vs. rare or uncommon), consistent with World Health Organization (WHO) guidelines. Data Collection Pathologists will independently review each case and provide diagnostic classifications along with confidence assessments using a 5-point scale. The digital pathology interface will automatically record time-to-diagnosis metrics. COSMO will process the same cases offline to generate independent diagnostic predictions and confidence scores. Both pathologist and AI predictions will be evaluated against established reference standard diagnoses. Analysis Framework The primary analysis will characterize diagnostic performance metrics (including accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and area under the receiver operating characteristic curve (AUROC)) for both pathologists (at the individual and aggregated levels) and the COSMO system. Secondary analyses will assess performance stratified by anatomical site, cancer incidence category, and pathologist experience level.