A recent study has reported of the success of an artificial intelligence (AI)-based algorithm in the detection, grading and clinical evaluation of prostate cancer from digitised slides.
Prostate cancer diagnosis
Adenocarcinoma of the prostate is the second most common cancer diagnosed in men, with more than 1 million new cases annually. To date, histopathological assessment of biopsy tissue is the mainstay diagnostic method for prostate cancer. It involves a core needle biopsy (CNB) and, if required, surgical resection. Most pathology laboratories currently perform light microscopic examination of haematoxylin and eosin (H&E)-stained tissue sections to make a diagnosis. Changing guidelines has led to an increase in the number of CNBs being reviewed. In addition, with the growing incidence of cancer cases and shortage of pathologists, the need for automated AI-based tools has become more apparent.
Management of males with prostate cancer relies on reliable diagnosis and prognosis (based on the Gleason grade). Therefore, adoption of automated tools will require clinical grade accuracy. Advancements in whole slide imaging have accelerated the adoption of digital imaging into pathology. As a result, there is now interest in utilising AI tools to analyse these digital images.
The aim of this study, published in The Lancet, was to clinically validate the AI-based algorithm in detecting and grading prostate adenocarcinoma. They also wanted to deploy this system in a pathology laboratory for clinical use.
Researchers developed the AI-based algorithm using digitised H&E-stained slides from prostate CNBs. The algorithm provided slide-level scores for probability of cancer, Gleason score 7-10 (high-grade cancers), Gleason pattern 5 (aggressive cancers), and perineural invasion and calculation of cancer percentage present in CNB material. The team validated the algorithm on an external dataset of 100 consecutive digitised cases of prostate CNB. They subsequently implemented the AI tool in a pathology laboratory within routine clinical workflow.
The algorithm achieved an AUC of 0.997 for cancer detection in the internal test set, and 0.991 in the external validation set. The AUC for distinguishing between a low-grade and high-grade cancer diagnosis was 0.941. In addition, the AUC for detecting Gleason pattern 5 was 0.871 in the external validation set. The team observed good agreement in the cancer percentages calculated by pathologists and the algorithm. The algorithm achieved an AUC of 0.957 for perineural invasion. When used in clinical practice, the algorithm assessed 11,429 H&E-stained slides (941 cases), leading to 90 Gleason score 7-10 alerts and 560 cancer alerts. The algorithm also detected a missed cancer that was originally diagnosed as benign.
This study reports the development of a medical-grade AI-based algorithm that can accurately evaluate digitised prostate CNB slides. It highlights the successful deployment of this AI in routine clinical practice. The algorithm shows high levels of accuracy in identifying and quantifying prostate cancer. It can also differentiate between graded tumours and detect perineural invasion in CNBs.
The team argue that deployment of such an AI tool in clinical practice is timely. Not only is this because of the prevalence of prostate cancer in men but is also due to the increasing workload for pathologists. There is an ongoing decline in the pathology workforce, which coincides with an increasing workload that could delay cancer diagnosis and cause diagnostic errors. Implementation of AI-based tools for routine clinical practice offer practical benefits. Consequently, this may lead to improved efficiency, accuracy, consistent diagnoses and improved patient management.