Medical Technology · global
AI-assisted prostate biopsy interpretation raises accuracy to 75% among less experienced pathology residents
A study of 15 residents suggests AI annotations may help less experienced readers identify lesions. But everyone interpreted the slides independently before viewing AI prompts, leaving the contribution of a second review to the improvement unresolved.
To identify cancer in prostate needle biopsies, pathology residents must make judgments based on subtle tissue morphology. Can interpretation become more accurate when AI highlights suspicious areas on the screen? A new study observed greater improvements among less experienced residents using assistance, but the findings leave open how much credit belongs to AI prompts and how much to taking another look.
The peer-reviewed study, published on November 4 in *BMC Medical Education*, invited 15 first- through third-year pathology residents to interpret 186 whole-slide digital images of prostate needle biopsies. Of these, 104 showed adenocarcinoma, 75 were benign, and another 7 showed atypical small acinar proliferation. The digital platform provided tumor region prompts and Gleason pattern annotations for residents to consult during interpretation.
The study grouped residents according to their previous experience in genitourinary pathology. Observed accuracy among those with less experience rose from 62.3% to 75.1%, an increase of approximately 13 percentage points; among those with more experience, it rose from 77.9% to 83.0%. Both groups improved with assisted interpretation, narrowing the initial gap. Scienmag's coverage of the same study also noted that the findings reflect performance while using the tool and cannot directly be regarded as evidence of training effectiveness.
The changes prompted by AI did not all lead toward correct answers. The study recorded that, of 786 initially incorrect interpretations, 307 became correct with assistance; conversely, of 1,913 initially correct interpretations, 75 became incorrect. These figures show that evaluating an assistive tool requires examining both how many errors it corrects and whether it leads readers away from initially correct judgments.
The most important limitation lies in the reading order: all residents first interpreted the slides independently, then reviewed the same slides with AI assistance. Because the order was fixed, the effects of a second review, including familiarity with the images and reconsideration, were intertwined with the effects of AI itself. Thus, although increased accuracy was observed in the study, it cannot be attributed entirely to AI.
The study also did not measure learning or retention, so it cannot answer whether residents would continue to interpret slides more accurately once the prompts were removed. Diagnostic performance on slides likewise cannot be directly translated into treatment or health outcomes for patients. The paper also disclosed that some authors were affiliated with NovinoAI, which developed the platform and AI models under evaluation; subsequent independent validation would help determine whether the findings can be generalized.
The study points to a concrete possibility: AI annotations may provide interpretation support in pathology training, particularly for those with limited experience. Confirming their independent benefit will require varying the reading order and tracking performance after assistance is removed, to clarify whether the tool provides immediate clues or also helps physicians develop lasting diagnostic skills.