Prediction of pathologic complete response to neoadjuvant chemotherapy in breast cancer (SWOG S0800) using image analysis-based tumor infiltrating lymphocyte measurements.

Abstract

594 Background: Image analysis-based tumor infiltrating lymphocyte (TIL) quantitation methods are being developed to eliminate reader-to-reader variation in TIL assessment that hinders its clinical adoption as prognostic and chemotherapy response predictive marker. We evaluated the ability of an image analysis-based TIL score to predict pathologic complete response (pCR) and event free survival (EFS) in breast cancer. Methods: 113 pretreatment samples were analyzed from the SWOG S0800 trial that randomized stage IIB-IIIC HER-2-negative breast cancers to neoadjuvant chemotherapy with or without bevacizumab. TIL quantitation was performed on H&E sections using QuPath open-source software and a convolutional neural network cell classifier (CNN11). The digital easTILs% score was calculated as [sum of TIL Area (mm 2 ) / Stromal Area (mm 2 )] x 100. Pathologist-read stromal TIL score (sTILs%) was defined using international guidelines. A previously validated threshold of easTILs% > 19.9% defined high easTILs% status. Results: Pretreatment easTILs% was significantly higher in cases with pCR compared to residual disease (RD) (means, 31% vs. 17%, p < 0.001). easTILs% high and low cases had pCR rates of 41% and 21% (p = 0.019), respectively. In logistic regression adjusting for other factors, easTILs% was prognostic for pCR as continuous score (p < 0.001) and as high vs low categories (p = 0.035). There was strong positive correlation between easTILs% and sTILs% (r = 0.606, p < 0.0001), and sTILs% was also predictive of pCR. The areas under the prediction curve (AUC) were 0.709 and 0.627 for easTILs% and sTILs%, respectively. There was no statistically significant interaction between easTILs% and bevacizumab benefit (p = 0.26), and higher easTILs% or sTILs were not associated with better EFS. Conclusions: Image analysis-based TIL quantification is predictive of pCR in breast cancer and had better pCR outcome discrimination than pathologist-read sTIL count. [Table: see text]