AI tool can detect prostate cancer missed by pathologists

An artificial intelligence (AI) model has demonstrated the ability to detect subtle morphological changes in tissue biopsies that indicate clinically significant prostate cancer before it becomes histologically apparent, according to data published in Nature.¹

Thus, the model was able to distinguish between those likely to develop clinically significant prostate cancer vs those expected to remain low-risk.

“The study has been nicknamed the ‘missed study,’ as the goal of finding the cancer was ‘missed’ by the pathologists,” explained senior author Carolina Wählby, PhD, of Uppsala University in Sweden, in a news release on the findings.² “We have now shown that with the help of AI, it is possible to find signs of prostate cancer that were not observed by pathologists in more than 80% of samples from men who later developed cancer.”

For the study, the investigators retrospectively assessed H&E-stained prostate needle biopsies from 232 men who had an elevated prostate-specific antigen (PSA) level and whose initial tissue samples were diagnosed as benign. The final cohort comprised 213 patients with 587 biopsy slides. All patients included in the analysis resided in the northern region of Sweden.

Among all 213 patients, 88 developed clinically significant prostate cancer within 30 months, and 125 had no cancer detected over 8 years. Patients were matched on age and PSA levels.

In patients who did develop clinically significant prostate cancer (n = 88), the AI model showed an area under the curve (AUC) of 0.81 and a sensitivity of 0.81 at a false positive rate of 0.26. At the patient level, the tool achieved an AUC of 0.82 and a sensitivity of 0.92 at a false positive rate of 0.32.

On these findings, the authors noted, “Considering these patients were not identified at all upon initial diagnosis, we consider the sensitivity compensates for the false positive rates.”

Specifically, the model highlighted patterns such as stromal collagen and altered glandular epithelial cells in those who did develop clinically significant prostate cancer. In the informative stroma areas, reduced numbers of smooth muscle cells and increased collagen/altered matrix were observed.

“When we looked at the patterns that the AI ranked as informative, we saw changes in the tissue surrounding the glands in the prostate – changes also observed in other studies,” Wählby added in the news release.² “This shows that AI analysis of routine biopsies can detect subtle signs indicating clinically significant prostate cancer before it becomes obvious to a pathologist.”

The investigators also found that the model was able to more easily identify patients who were later diagnosed as ISUP1 as “low risk” vs those who remained benign. The tool also showed a difference in performance “in identifying patients who were subsequently diagnosed with high grade [prostate cancer] (ISUP4-5) and patient at intermediate risk such as ISUP 2 or ISUP 3,” the authors noted.

Overall, the authors suggest that a tool such as this could be used to augment clinical decision-making, helping to determine how soon men identified as having benign lesions should be followed up.

“If these findings were confirmed, the methodology could be used as a complementary screening step,” they conclude.

REFERENCES

1. Chelebian E, Avenel C, Järemo H, et al. Discovery of tumour indicating morphological changes in benign prostate biopsies through AI. Sci Rep. 2025;15(1):30770. doi:10.1038/s41598-025-15105-6

2. AI can find cancer pathologists miss. News release. Uppsala University. August 22, 2025. Accessed August 26, 2025. https://www.eurekalert.org/news-releases/1095642