New AI Research Looks to Detect Prostate Cancer
Researchers with the Emory University Empathetic AI for Health Institute authored the article, “Stress testing deep learning models for prostate cancer detection on biopsies and surgical specimens,” recently published in the Journal of Pathology.
The study analyzed whether artificial deep-learning models could detect prostate cancer when applied to different types of histological samples from biopsies (BX) and radical prostatectomies (RP).
AI models were tasked with identifying structural differences between BX and RP samples, focusing on gland structure variations. Using a dataset of more than 1,000 samples, the study shows that models trained on one sample type perform less accurately on the other.
The results highlight that BX-trained models excel at identifying small, closed gland structures but struggle with large, open glandular structures typical of RP samples. Conversely, RP-trained models perform well on open gland structures but have difficulty with smaller, BX-specific features. A model trained on both sample types achieved average performance but did not significantly improve accuracy for either type individually.
The study concludes that cancer detection models should be tailored to the specific sample type (BX or RP) to ensure accuracy. It advocates against using BX-trained models for RP slides and vice versa in clinical practice, as morphological differences impact predictions.
To view the research in its entirety, click the link below below:
https://pathsocjournals.onlinelibrary.wiley.com/doi/full/10.1002/path.6373