Machine learning for rhabdomyosarcoma histopathology.

Area of Science:

• Oncology
• Pathology
• Artificial Intelligence

Background:

• Accurate diagnosis of pediatric soft-tissue sarcomas is crucial for effective treatment.
• Limited access to specialized sarcoma pathologists globally, particularly in low-resource settings, hinders timely diagnosis.
• Developing accessible diagnostic tools is essential to bridge this expertise gap.

Purpose of the Study:

• To develop and validate a deep learning-based diagnostic system for soft-tissue sarcoma subtypes.
• To create a convolutional neural network (CNN) tool that assists pathologists by quantifying diagnosis likelihood from histopathology slides.
• To improve early differential diagnosis of pediatric/young adult sarcomas.

Main Methods:

• A CNN model was trained on 424 whole histopathology slides of alveolar rhabdomyosarcoma, embryonal rhabdomyosarcoma, and clear-cell sarcoma.
• The model's performance was evaluated on withheld testing cohorts, achieving an area under the receiver operating characteristic curve (AUC) above 0.889 for all subtypes.
• External validation included human and mouse model sarcoma samples, assessing robustness against variations like anaplasia.

Main Results:

• The CNN model demonstrated high accuracy in classifying sarcoma subtypes, with AUC values exceeding 0.889.
• The system successfully classified external human and mouse-derived sarcoma samples.
• The model showed robustness when tested against histopathological variations and untrained disease models.

Conclusions:

• The developed CNN-based system shows significant potential as a pre-pathologist screening tool for soft-tissue sarcomas.
• Machine learning can assist local pathologists in rapidly narrowing differential diagnoses for pediatric, adolescent, and young adult sarcomas.
• This technology can enhance diagnostic capabilities in regions with limited access to specialized pathology expertise.