An interpretable deep learning model for predicting endometrial cancer molecular subtypes from H&E-stained slides

Area of Science:

• Oncology
• Computational Pathology
• Artificial Intelligence in Medicine

Background:

• Endometrial cancer molecular subtypes are crucial for prognosis and treatment.
• Accurate prediction of these subtypes is essential for personalized medicine.

Purpose of the Study:

• To develop an interpretable deep learning model for predicting endometrial cancer molecular subtypes from H&E-stained whole slide images (WSIs).
• To validate the model's generalizability and clinical applicability across different cohorts.
• To correlate histological features with molecular subtypes at macro- and micro-levels.

Main Methods:

• Training an end-to-end deep learning network on the Fudan cohort (n=364) for molecular subtype prediction.
• Validating the model using external cohorts: TCGA (n=296) and Suzhou (n=36).
• Assessing model performance using area under the receiver operating characteristic curve (AUROC) and analyzing morphological features.

Main Results:

• The model achieved a macro-average AUROC of 0.867 in cross-validation.
• Class-wise AUROCs ranged from 0.835 to 0.910 for different subtypes (MSI-H, NSMP, p53abn, POLEmut).
• Distinct morphological features were identified for each subtype, including stromal lymphocytic infiltration (MSI-H), heterogeneity (POLEmut), papillary growth (p53abn), and stromal cellularity (NSMP).

Conclusions:

• The developed deep learning model offers an accurate and interpretable method for predicting endometrial cancer molecular subtypes from WSIs.
• The findings provide a theoretical basis for utilizing histological features for molecular subtype prediction and guiding individualized treatment strategies.
• This approach has potential clinical applicability for non-invasively determining molecular subtypes, aiding in treatment decisions.