Deep learning algorithms from histopathological images stratify molecular subtypes for leiomyosarcoma: a...

Area of Science:

• Computational pathology
• Oncology
• Artificial intelligence in medicine

Background:

• Leiomyosarcoma (LMS) is the most common soft tissue sarcoma.
• Molecular subtypes of LMS are linked to prognosis and therapeutic sensitivity.
• Accurate subtyping and survival prediction are crucial for LMS patient management.

Purpose of the Study:

• To develop and validate deep learning (DL) algorithms for LMS molecular subtyping.
• To develop and validate DL algorithms for predicting LMS survival.
• To assess the utility of DL-derived histomorphological features in assisting pathologists.

Main Methods:

• DL models were trained on whole slide images (WSIs) from The Cancer Genome Atlas (TCGA).
• External validation was performed on a multi-center, real-world cohort.
• Performance metrics included AUROC, accuracy, recall, specificity, precision, and F1 score. Computation visualizations (CVs) were used to interpret DL models.

Main Results:

• The LMS-DL model achieved an AUROC of 0.944 for molecular subtyping and 0.937 for predicting two-year overall survival.
• The DL model outperformed pathologists in subtyping accuracy by over 30%.
• CVs improved pathologists' subtyping accuracy by 12.1% with reduced diagnostic time and uncertainty.

Conclusions:

• DL models show high accuracy and efficacy in LMS molecular subtyping and survival prediction.
• These models can significantly reduce pathologists' workload.
• Further validation in large prospective cohorts is warranted.