(P 123) ARTIFICIAL INTELLIGENCE-POWERED MITOSIS EVALUATION TO COMPLEMENT HUMAN MITOTIC RATE INTERPRETATION IN PREDICTING PROGNOSIS IN LIPOSARCOMA

Objective: Liposarcoma is one of the most common adult sarcomas, with prognosis varying between well-differentiated (WD) and dedifferentiated (DD) types. Distinguishing between WD and DD liposarcoma can be challenging, particularly when the low-grade (LG) spindle cell component does not meet the criteria for typical DD liposarcoma. Recently, the mitotic rate (MR) has been recognized as a significant prognostic factor, with at least 5 mitotic figures per 10 high-power fields (HPFs) suggested as the diagnostic criterion for DD liposarcoma. However, assessing MR by a pathologist’s choice of 10 HPFs may not accurately reflect the overall MR on the whole slide/tumor and may have inter-rater discrepancies. Artificial intelligence (AI) analyzers, including those assessing MR, have recently been used as complementary tools to improve pathologists’ interpretations in various diagnostic situations.

Methods: Lunit SCOPE IO was trained with 22,083 mitotic cells from 527 H&E whole slide images (WSI) spanning a pan-cancer dataset of more than 20 cancer types. The model is capable of detecting WSI-level mitosis and calculating mitotic cell density. It was applied to 89 surgically resected liposarcoma cases from Seoul National University Bundang Hospital to validate the model's performance. The dataset consisted of 74 DD types and 15 WD types. All slides were evaluated for MR by board-certified pathologists, defining more than 5 mitoses per 10 HPFs (2mm²) as a high-risk group. Clinical outcome was assessed by recurrence-free survival (RFS).

Results: The pathologist-assessed MR averaged 11.39 ± 13.91 per 10 HPFs (mean ± standard deviation), while the AI-MR averaged 2.85 ± 4.98 per mm², which is equivalent to 5.71 ± 9.97 per 10 HPFs. The Pearson correlation value between the two results was 0.497. When the high-risk group was classified based on the pathologist's MR, 52 cases (58.4%) were categorized into that group. For AI-MR, the cutoff was set to 0.80 per mm² to divide the high-risk group into equal proportions (58.4%). The high-risk group based on pathologist readings had a hazard ratio (HR) of 0.18 (95% confidence interval [CI] 0.08–0.41) compared to the low-risk group, while the high-risk group based on AI-MR had an HR of 0.26 (95% CI 0.12–0.58) compared to the low-risk group. Relative to cases with both human and AI assessments categorized in the low-risk group (reference), cases classified as high-risk by both had a significantly higher risk of recurrence, with an HR of 6.64 (95% CI 2.55–17.3). The HR was 2.41 (95% CI 0.57–10.2) for the human-only high-risk group, which was not significant, and the AI-only high-risk group showed no significant difference compared to the reference.

Conclusion: This study suggests that AI-MR analysis at the WSI level may serve as a valuable and unexpected complement to traditional HPF-based MR evaluations by pathologists in liposarcoma cases.