The Application Of Ultrasound Features And Ultrasound Based Deep Learning Models In Evaluating TILs Levels In Breast Cancer

Purpose: Tumor-infiltrating lymphocytes(TILs)have emerged as an efficient biomarker predicating treatment response and prognosis of breast cancer.This study aimed to evaluate the association between conventional ultrasound and contrast-enhanced ultrasound(CEUS)imaging features with TILs levels in invasive breast cancer patients,explore the feasibility of using a deep-learning(DL)approach to predict TILs levels in breast cancer from ultrasound(US)images,and screen the optimal deep learning model to predict TILs level of breast cancer.Than we evaluated the feasibility of TILs-deep learning score(DLS)as a substitute for TILs levels assessed by biopsy in predicting the efficacy of neoadjuvant chemotherapy(NAC)in patients with HER2-positive and triple-negative breast cancer.Methods: 1.We retrospectively included 267 women with invasive BC who had undergone conventional ultrasound and CEUS.According to pathological findings,patients were divided into low TILs group(≤10%)and high TILs group(>10%).Logistics regression analysis was used to screen independent predictors of breast cancer TILs level,and multiple regression models were established.2.A total of 494 breast cancer patients with pathologically confirmed invasive breast cancer from two hospitals were retrospectively enrolled.396 cases from hospital 1 were divided into training cohort(n = 298)and internal validation cohort(n = 98),and patients from hospital 2(n = 98)were used as external validation cohort.Five different deep learning frameworks were trained with ultrasound images of the training cohort,so as to construct and screen the deep learning model that can predict the optimal level of TILs in breast cancer.3.Eighty-six pathologically confirmed patients with invasive breast cancer who were expected to receive neoadjuvant chemotherapy were retrospectively included,with molecular subtype of HER2-positive(n = 62)and triple-negative(n = 24).Ultrasound images of pretreatment breast cancer were input to train attention-Dense Net121 model after image preprocessing,and TILs-DLS of each case image were output.According to the pathological results,patients were divided into effective group(n=48)and ineffective group(n=38).The clinicpathological characteristics of patients,TILs-DLS and TILs pathological score were included in Logistics regression analysis to screen independent predictors of NAC efficacy.The consistency test was used to compare the consistency between TILs-DLS and TILs pathological scores,and the kappa coefficient was calculated.Delong test was used to compare the difference between TIls-DLS and TILs pathological scores in predicting the efficacy of neoadjuvant chemotherapy.Results: 1.There were significant differences between the high TILs group and the low TILs group in shape,margin,posterior echo,enhanced homogeneity,PI value and molecular subtype.In the high TILs group,the lesions were mainly oval or round in shape,with clear margin and posterior enhancemrnt.In contrast,the lesions in the low TILs group were more likely to show irregular shape,unclear margin,and attenuation of echo behind the lesions.In the qualitative and quantitative parameters of contrast-enhanced ultrasound,compared with low TILs,high TILs lesions were more likely to show regular shape,clear margin,homogeneous enhancement and higher peak intensity(PI)than low TILs lesions(all P < 0.05).Multiple Logistic regression analysis showed that lesion shape,posterior echo,PI value and enhanced homogeneity were independent predictors of breast cancer TILs level.The multiple regression model combined with the above four independent predictors showed moderate performance in predicting breast cancer TILs level,with an AUC of 0.79 and a sensitivity of 72.4%.The specificity was 77.9%.2.The five deep learning models based on ultrasound images can significantly distinguish between high and low TILs lesions in the training and validation cohort.Attention-Dense Net121 has the best overall performance,with AUC of 0.922,F1 score of 0.830,accuracy of 79.5%,sensitivity of 90.7%,specificity of 65.9%,positive predictive value of 76.6%,and negative predictive value of 85.3% in the internal verification set.In the external validation cohort,AUC was 0.873,F1 score was 0.851,accuracy was 83.6%,sensitivity was 85.2%,specificity was 81.8%,positive predictive value was 85.2%,and negative predictive value was 81.8%.In addition,external validation focused on the attention-Dense Net121 model’s ability to distinguish TILs well in stratified analyses of each molecular subtype of breast cancer.3.Univariate and multivariate logistic regression analysis showed that lesion size,TILs pathological score and TILs-DLS were independent predictors of the efficacy of neoadjuvant chemotherapy.In terms of predicting the efficacy of neoadjuvant chemotherapy,the diagnostic performance of TILs-DLS(AUC= 0.811)was similar to that of TILs pathological score(AUC=0.806),with no statistical significance(P = 0.922).Meanwhile,there was a high consistency between TILs-DLS and TILs pathological score in predicting the efficacy of NAC,with a kappa coefficient of 0.61.Conclusions: Conventional ultrasound and CEUS features were associated with TILs levels in invasive breast cancer.The deep learning model based on ultrasound imaging of breast cancer has high accuracy,sensitivity and specificity,which can distinguish the level of breast cancer TILs and perform well in hierarchical analysis of different molecular types of breast cancer.Meanwhile,the diagnostic performance of TIL-DLS and TILs is similar,and the two are highly consistent.TILs-DLS is expected to be a new method for non-invasive assessment of TILs levels,and aid in individualized treatment decision making.