Predicting Prostate Cancer Molecular Subtype With Deep Learning on Histopathologic Images

Area of Science:

• Computational pathology
• Genomic medicine
• Artificial intelligence in oncology

Background:

• Morphologic features in prostate cancer lack reproducible associations with molecular alterations.
• Hematoxylin and eosin (H&E)-stained whole slide images (WSI) offer potential for deep learning algorithms to detect genomic alterations.

Purpose of the Study:

• To develop and validate deep-learning algorithms for identifying prostate tumors with ETS-related gene (ERG) fusions or PTEN deletions using H&E-stained WSI.
• To assess the performance of these algorithms in predicting genomic status across radical prostatectomy and needle biopsy cohorts.

Main Methods:

• A novel transformer-based hierarchical architecture was employed for automated tumor identification, feature representation learning, classification, and explainability map generation.
• Two distinct vision transformer-based networks and a transformer-based classifier were trained on WSI from radical prostatectomy cohorts with known ERG/PTEN status.
• Algorithm performance was validated across multiple independent radical prostatectomy and needle biopsy cohorts.

Main Results:

• The ERG algorithm demonstrated strong performance across validation cohorts, with AUCs ranging from 0.86 to 0.91 for radical prostatectomy and 0.78 to 0.80 for needle biopsies.
• The PTEN algorithm achieved AUCs ranging from 0.72 to 0.81 for radical prostatectomy and 0.75 for needle biopsies (focusing on homogeneous PTEN status).
• For heterogeneous PTEN loss, the algorithm's predicted PTEN loss percentage correlated significantly with immunohistochemistry (r=0.58, P=0.0097).

Conclusions:

• Deep-learning algorithms can accurately predict ERG fusions and PTEN deletions in prostate cancer from H&E-stained WSI.
• These algorithms show potential for screening clinically relevant genomic alterations non-invasively.
• H&E images, analyzed by AI, can serve as a tool to identify underlying genomic alterations in prostate cancer.