Deep Learning of Rhabdomyosarcoma Pathology Images for Classification and Survival Outcome Prediction

Area of Science:

• Computational pathology
• Pediatric oncology
• Biomarker discovery

Background:

• Rhabdomyosarcoma (RMS) is the most common pediatric soft tissue sarcoma.
• Histologic subtypes of RMS impact treatment and prognosis.
• Accurate classification and prognostic prediction are crucial for patient management.

Purpose of the Study:

• To develop computational algorithms for RMS histology subtype classification.
• To create a prognostic model for embryonal rhabdomyosarcoma (eRMS) risk stratification.
• To leverage AI for improved diagnostic and prognostic capabilities in RMS.

Main Methods:

• Development of a convolutional neural network (CNN) for RMS histology classification using digitized pathology images.
• Creation of a prognostic model for eRMS utilizing patient data.
• Validation of models on independent test datasets.

Main Results:

• The RMS classification model achieved high performance (AUC 0.94 for alveolar RMS, 0.92 for eRMS).
• The eRMS prognostic model identified distinct high- and low-risk groups with significantly different event-free survival (P=0.02).
• Predicted risk group remained a significant independent predictor of event-free survival (HR 4.64, P=0.04).

Conclusions:

• This study presents the first computational algorithms for RMS histopathology classification and prognosis.
• AI-driven models can assist pathologists in classification and provide valuable parameters for risk stratification.
• These tools have the potential to enhance patient care for children with rhabdomyosarcoma.