New AI System Cuts Through Lab Bias in Prostate Cancer Diagnosis
Researchers have developed an AI system that diagnoses prostate cancer more reliably across different hospitals and labs than existing approaches. The breakthrough trains on patient outcomes rather than subjective expert opinions, potentially reducing misdiagnosis rates and standardizing care across healthcare systems.
Originaltitel: Robust, credible, and interpretable AI-based histopathological prostate cancer grading
<p>Background: Prostate cancer (PCa) is among the most common cancers in men and itsdiagnosis requires the histopathological evaluation of biopsies by human experts. Whileseveral recent artificial intelligence-based (AI) approaches have reached human expert-levelPCa grading, they often display significantly reduced performance on external datasets. Thisreduced performance can be caused by variations in sample preparation, for instance thestaining protocol, section thickness, or scanner used. Another limiting factor of contemporaryAI-based PCa grading is the prediction of ISUP grades, which leads to the perpetuation ofhuman annotation errors.</p><p>Methods: We developed the prostate cancer aggressiveness index (PCAI), an AI-based PCadetection and grading framework that is trained on objective patient outcome, rather thansubjective ISUP grades. We designed PCAI as a clinical application, containing algorithmicmodules that offer robustness to data variation, medical interpretability, and a measure ofprediction confidence. To train and evaluate PCAI, we generated a multicentric, retrospective,observational trial consisting of six cohorts with 25,591 patients, 83,864 images, and 5 yearsof median follow-up from 5 different centers and 3 countries. This includes a high-variancedataset of 8,157 patients and 28,236 images with variations in sample thickness, stainingprotocol, and scanner, allowing for the systematic evaluation and optimization of modelrobustness to data variation. The performance of PCAI was assessed on three external testcohorts from two countries, comprising 2,255 patients and 9,437 images.</p><p>Findings: Using our high-variance datasets, we show how differences in sample processing,particularly slide thickness and staining time, significantly reduce the performance ofAI-based PCa grading by up to 6.2 percentage points in the concordance index (C-index). Weshow how a select set of algorithmic improvements, including domain adversarial training,conferred robustness to data variation, interpretability, and a measure of credibility to PCAI.These changes lead to significant prediction improvement across two biopsy cohorts and oneTMA cohort, systematically exceeding expert ISUP grading in C-index and AUROC by up to22 percentage points.</p><p>Interpretation: Data variation poses serious risks for AI-based histopathological PCagrading, even when models are trained on large datasets. Algorithmic improvements formodel robustness, interpretability, credibility, and training on high-variance data as well asoutcome-based severity prediction gives rise to robust models with above ISUP-level PCagrading performance.</p>