AI improves physician accuracy in identifying prostate cancer extent

Use of artificial intelligence (AI)-assisted cancer contours helps minimize underestimation of the extent of prostate cancer, thereby improving the physician’s ability in terms of contouring accuracy and negative margin rate, a study has shown. 

This technology has the potential to enhance clinical outcomes, given that accurate contouring “informs patient management strategy and underpins the oncologic efficacy of treatment,” according to the investigators, noting that AI-assisted cancer contouring “overcomes many limitations of the current clinical workflow,” which then allows “patient-specific therapy selection and treatment planning.” 

In this study, seven urologists and three radiologists from five institutions, with experience ranging from 2 to 23 years, interpreted the prostatectomy cases. Each clinician assessed 50 cases retrospectively eligible for focal therapy, and each case had a T2-weighted MRI, contours of the prostate and region/s of interest suspicious for cancer, as well as a biopsy report. 

Readers first defined cancer contours cognitively by manually delineating tumour boundaries to capture all clinically significant diseases. After ≥4 weeks, they contoured the same cases using AI software. The investigators then assessed the AI-assisted, cognitively defined, and hemigland cancer contours using tumour boundaries on whole-mount histopathology slides as ground truth. 

Accuracy and negative margin rate of cancer contours served as the primary outcomes. Generalized estimating equations were used to perform all statistical analyses. 

AI-assisted cancer contours demonstrated superior balanced accuracy, defined as a mean of voxel-wise sensitivity and specificity relative to cognitively defined and hemigland contours (84.7 percent vs 67.2 percent and 75.9 percent, respectively; p<0.0001). [J Urol 2024;212:52-62] 

Notably, cognitively defined cancer contours grossly underestimated the extent of prostate cancer, as shown by a negative margin rate of 1.6 percent as opposed to 72.8 percent for AI-assisted cancer contours (p<0.0001). 

“Defining prostate cancer contours is a complex task, undermining the efficacy of interventions such as focal therapy,” the investigators said. 

“Given access to standard clinical information, physicians tended to err toward small ‘high-specificity’ contours that fail to encapsulate the entire tumour,” they noted. This tendency usually occurs since many physicians rely on the MRI appearance of a tumour. However, the real extent of prostate cancer is often “MRI-invisible.” 

Underestimation 

Previous studies have reported the underestimation of tumour size, which is potentially responsible for the shifting rates of residual disease reported during focal treatment trials. [J Urol 2017;197:320-326; J Urol 2021;205:444-451; Eur Urol Open Sci 2023;54:20-27; Cancer Med 2023;12:9351-9362] 

“AI can help resolve this challenging problem,” according to the investigators, noting how “AI was significantly more accurate than hemigland and standard-of-care (SOC) contours, indicating a better balance between sensitivity and specificity.”  

In addition, the use of AI influenced clinical decision-making, with changes in recommended treatment seen in more than a quarter of cases. 

“The cancer estimation map improved confidence in recommending focal therapy over whole gland therapy, with urologists trending towards a more targeted approach,” the investigators said.  

“Clinician reluctance to recommend focal therapy under SOC conditions mirrors documented shortcomings in patient selection based on MRI and systematic biopsy alone,” they added. [Eur Urol 2019;75:712-720; Cancer 2019;125:2955-2964]