Development and deployment of a histopathology-based deep learning algorithm for patient prescreening in a clinical...

Area of Science:

• Oncology
• Pathology
• Artificial Intelligence

Background:

• Accurate identification of genetic alterations like Fibroblast Growth Factor Receptor is vital for targeted cancer therapies.
• Current molecular testing methods can be time-consuming and tissue-intensive, potentially delaying patient care and clinical trial enrollment.
• Developing AI-based solutions for biomarker detection could streamline these processes.

Purpose of the Study:

• To develop and validate a deep-learning algorithm for detecting biomarkers from H&E-stained whole slide images in advanced urothelial cancers.
• To assess the algorithm's potential to reduce the need for molecular testing and accelerate patient recruitment for clinical trials.
• To evaluate the real-world applicability and cost-saving potential of the AI system in clinical settings.

Main Methods:

• A deep-learning algorithm was developed using over 3000 H&E-stained whole slide images from patients with advanced urothelial cancers.
• The algorithm was optimized for high sensitivity to minimize the exclusion of eligible patients.
• Validation was performed on an independent dataset of 350 patients, and the system was deployed across 89 global clinical sites.

Main Results:

• The algorithm achieved an area under the curve (AUC) of 0.75.
• It demonstrated a specificity of 31.8% at 88.7% sensitivity.
• A projected 28.7% reduction in molecular testing was estimated, with successful deployment in a multi-site study.

Conclusions:

• The developed AI algorithm shows promise in accurately identifying potential biomarker status from H&E images.
• This approach can significantly reduce reliance on molecular testing, leading to cost savings and faster patient access to targeted therapies.
• The successful deployment indicates the system's potential to optimize resource allocation in drug development and clinical practice.