Systematic PD-L1 Slide Analysis Based on Multi-Objective Learning
In treatment of cancers,especially non-small-cell lung cancers such as lung squamous cell carcinoma(LUSC),tumor proportion score(TPS)of a programmed death-ligand 1(PD-L1)slide is essential for selecting tumor therapies.Many parameters of tumor cells(TCs)are vital to cancer diagnosis.Although the indexes can be estimated via the computational analysis,there is seldom a unified system that could acquire different nucleus information simultaneously.To address the issues,multi-objective learning pipeline(MOLP)is proposed to predict TPS,cell counts,nucleus contours and categories altogether from PD-L1 slides of LUSC.The main network comprises two branches,one estimating TPS via the cell analysis and the other directly regressing TPS.It minimizes the difference between these two approximated values of TPS to gain robustness.The cell-analysis branch increases confidence of the estimated TPS by nucleus segmentation,classification and counting.It also enables the system to estimate appearance parameters of TCs for LUSC diagnosis.Experiments on a large image set show that MOLP is feasible and effective.The TPS predicted by MOLP exhibits statistically significant correlation with pathologists'scores,with a mean absolute error(MAE)of 4.97(95％confidence interval(CI):-0.56-10.49)and a Pearson correlation coefficient(PCC)of 0.97(p＜0.001).
国家科技期刊平台
NSTL
万方数据
