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

Area of Science:

• Computational pathology
• Artificial intelligence in oncology
• Digital histopathology

Background:

• Pediatric sarcomas are rare and diagnostically challenging, often needing specialized expertise and expensive genetic tests.
• Current diagnostic methods for pediatric sarcomas can be limited by rarity and subtype diversity.
• Overcoming diagnostic barriers in pediatric sarcoma classification is crucial for timely and accurate treatment.

Purpose of the Study:

• To develop and validate a deep learning pipeline for accurate pediatric sarcoma subtype classification using digitized histology slides.
• To assess the generalizability of computational models across multiple institutions and datasets.
• To improve the efficiency and accessibility of pediatric sarcoma diagnosis through advanced AI methods.

Main Methods:

• A harmonized dataset of 867 whole slide images (WSIs) from three medical centers and the Children's Oncology Group (COG) was utilized.
• Multiple deep learning architectures, including convolutional neural networks (CNNs) and vision transformers (ViTs), were evaluated.
• Input parameters like tile size and resolution were optimized, and SAMPLER-based WSI representations were employed.

Main Results:

• Advanced ViT foundation models (UNI, CONCH) significantly outperformed previous methods.
• The pipeline achieved high performance in distinguishing rhabdomyosarcoma (RMS) from non-rhabdomyosarcoma (NRSTS) (AUC 0.969) and RMS subtypes (AUC 0.961).
• A two-stage approach successfully identified Ewing sarcoma from other NRSTS (AUC 0.929) with significantly faster training times compared to conventional transformers.

Conclusions:

• Digital histopathology combined with rigorous image harmonization offers a powerful solution for pediatric sarcoma classification.
• Deep learning models, particularly advanced ViTs, demonstrate high accuracy and efficiency in classifying rare pediatric cancers.
• This computational approach has the potential to reduce diagnostic challenges and improve patient outcomes for pediatric sarcoma.