Clinicopathological Characterization of Pediatric Atypical Teratoid/Rhabdoid Tumors and an HE-IHC Dual-Path Deep Learning Model for Auxiliary Diagnosis.
AI interpretation is pending for this paper.
Open original publication →What the AI sees
Not AI summarized yet.
Research significance
Pending deeper interpretation.
Source abstract
Background/Objectives: Atypical teratoid/rhabdoid tumor (AT/RT) is a rare and aggressive pediatric embryonal tumor of the central nervous system with marked histological and immunophenotypic heterogeneity, which can make diagnosis difficult in some cases. This study aimed to summarize the clinicopathological and molecular features of pediatric AT/RT and to evaluate an HE-IHC dual-path deep learning model as an auxiliary diagnostic approach. Methods: Clinical, histopathological, immunophenotypic, ultrastructural, and fluorescence in situ hybridization (FISH) data were retrospectively collected from 18 children with AT/RT treated at Beijing Children's Hospital between February 2010 and April 2021. A total of 361 pathological images were used to train and test a ResNet50-based dual-path classification model with transfer learning and feature fusion. An additional independent test set of 175 histological and immunohistochemical images from six newly collected patients was used for supplementary validation. Results: The mean age at diagnosis was 2 years and 3 months. All cases showed loss of INI1 expression, positivity for CK and EMA, and a high Ki-67 index. FISH analysis identified SMARCB1 deletion in 7 of 15 tested cases. In the original image-based test set, the dual-path model achieved an accuracy of 90.91%, compared with 81.82% for the model without transfer learning, 86.36% for the single-path immunohistochemistry model, and 50.00% for the single-path histological model. In the additional independent test set, the trained model correctly classified all 175 images. Conclusions: Pediatric AT/RT shows diverse clinicopathological features and complex SMARCB1 alteration patterns. The HE-IHC dual-path model showed encouraging preliminary performance for auxiliary pathological assessment; however, larger multicenter cohorts with molecular subgroup annotation are needed for further validation before routine clinical application.