Abstract
Abstract:
Merkel cell carcinoma (MCC) is a rare, aggressive neuroendocrine cutaneous malignancy that can arise through 2 oncogenic pathways: a Merkel cell polyomavirus–positive pathway and a virus-negative pathway driven by ultraviolet exposure. Cytokeratin 20 (CK20), expressed in a paranuclear dot-like pattern, is a cornerstone diagnostic marker for MCC, yet roughly 5% of cases are CK20 negative, an uncommon immunophenotype that can cause diagnostic difficulty. In this exploratory study, we asked whether CK20-negative tumors show a distinct nuclear phenotype on digital image analysis. Using the HALO artificial intelligence platform (Indica Labs), we performed artificial intelligence–assisted whole-slide image analysis of 5 CK20-negative and 36 CK20-positive MCC sections (from 30 patients), extracting nuclear and cytoplasmic metrics from manually annotated tumor regions. Because sections from the same patient are not independent, we repeated the primary comparison at the patient level across all 24 possible one-section-per-patient combinations. CK20-negative MCC showed lower average nucleus roundness than CK20-positive tumors; the difference was significant by the Welch
t
test in the section-level analysis (0.731 vs. 0.754;
P
= 0.012) and remained significant by the Welch
t
test across all 24 patient-level combinations (
P
= 0.032 to 0.048), although the Mann–Whitney
U
test was no longer significant at the patient level. We also observed trends toward smaller nuclear area and perimeter, while cell, cytoplasmic, and tissue areas did not differ. Because Merkel cell polyomavirus status was unavailable, these differences cannot be attributed specifically to CK20 or viral status. These preliminary findings suggest that computational pathology can detect nuclear differences in this rare subset; larger, molecularly characterized cohorts from multiple institutions are needed for validation.