Multicenter Histology Image Integration and Multiscale Deep Learning for Machine Learning-Enabled Pediatric Sarcoma...

Area of Science:

• Computational pathology
• Artificial intelligence in oncology
• Digital histopathology

Background:

• Pediatric sarcomas are rare and diverse, posing diagnostic challenges that require specialized expertise and expensive genetic tests.
• Existing diagnostic methods can be limited by inter-observer variability and accessibility issues.

Purpose of the Study:

• To develop and validate a deep learning computational pipeline for accurate classification of pediatric sarcoma subtypes using digitized histology slides.
• To overcome diagnostic barriers by creating a generalizable and efficient classification system.

Main Methods:

• A harmonized dataset of 867 whole slide images (WSIs) from multiple institutions was used.
• Various deep learning architectures, including convolutional neural networks (CNNs) and vision transformers (ViTs), were evaluated.
• Input parameters like tile size and resolution were optimized for feature extraction using SAMPLER-based WSI representations.

Main Results:

• Advanced vision transformer (ViT) foundation models (UNI, CONCH) demonstrated superior performance.
• The models achieved high AUC values: 0.969±0.026 for distinguishing rhabdomyosarcoma (RMS) from non-rhabdomyosarcoma (NRSTS), and 0.961±0.021 for differentiating RMS subtypes.
• A two-stage pipeline identified Ewing sarcoma with an AUC of 0.929.
• SAMPLER-based classifiers were significantly more lightweight and faster to train than conventional transformer architectures.

Conclusions:

• Digital histopathology combined with rigorous image harmonization offers a powerful solution for pediatric sarcoma classification.
• The developed models reduce inter-observer variability and enhance diagnostic precision.
• This approach has the potential to improve global accessibility to diagnostics, leading to faster treatment planning for pediatric sarcoma patients.