Deep learning-based histologic classifiers enable molecular subtyping of metastatic prostate cancer

Area of Science:

• Computational pathology
• Artificial intelligence in oncology
• Prostate cancer research

Background:

• Metastatic prostate cancer is heterogeneous, often developing resistance to therapy.
• Neuroendocrine prostate cancer (NEPC) presents a challenging phenotype with poor prognosis.
• Accurate NEPC diagnosis is crucial but hindered by its complexity.

Purpose of the Study:

• To develop a deep learning model for diagnosing NEPC using H&E stained tissue images.
• To assess the model's ability to provide prognostic information for advanced prostate cancer.

Main Methods:

• Development of NEURAL-PC, a deep learning model using interpretable cellular features.
• Application of a multiple instance learning (MIL) framework for image classification.
• Validation of the model on independent external datasets.

Main Results:

• NEURAL-PC achieved an AUROC of 0.921 for NEPC classification in external validation.
• The model demonstrated robust diagnostic capabilities using only H&E images.
• NEURAL-PC provided prognostic information, enabling subclassification of advanced prostate cancer.

Conclusions:

• NEURAL-PC offers a reliable method for diagnosing NEPC from routine pathology slides.
• The model shows strong generalizability and prognostic value across diverse datasets.
• This hybrid deep learning approach advances diagnostic and prognostic tools in prostate cancer pathology.