1、Survival Benefit With IMRT Following Narrow-Margin Hepatectomy In Patients With Hepatocellular Carcinoma 2、Phenotypic Subgroups Of Early-Stage (T1-2NO) Breast Cancer And Individually Prediction Of Disease Outcomes Based On Deep Learning

Part I:Survival benefit with IMRT following narrow-margin hepatectomy in patients with hepatocellular carcinoma close to major vesselsObjective:The aim of this study was to investigate the role of post-operative intensity-modulated radiotherapy (IMRT) in patients receiving narrow-margin hepatectomy for hepatocellular carcinoma (HCC) located close to the major vessels.Material and methods:This exploratory study involved 181 HCC patients. Of them, 116 were treated with narrow-margin (< 1.0 cm) hepatectomy. Thirty-three of the 116 underwent postoperative IMRT (Group A), while 83 did not receive radiotherapy (Group B). The remaining 65 patients underwent wide-margin (>1.0 cm) hepatectomy (Group C). Prognosis and patterns of recurrence were assessed in the three groups.Results:The 3-year overall survival (OS) and disease-free survival (DFS) rates were 89.1 and 64.2% in Group A,67.7 and 52.2% in Group B and 86.0 and 60.1% in Group C respectively. The OS and DFS of Group A and Group C patients surpassed those of Group B patients (Group A vs. B, P = 0.009 and P = 0.038; and Group C vs. B, P = 0.002 and P = 0.010). Patients in Groups A and C experienced significantly fewer early recurrences than did patients in Group B (P = 0.002). Furthermore, patients in Groups A and C experienced substantially fewer intrahepatic marginal (P = 0.048) and diffuse recurrences (P = 0.018) and extrahepatic metastases (P = 0.038) than did patients in Group B. No patient developed radiation-induced liver disease.Conclusion:The postoperative IMRT following narrow-margin hepatectomy to be a favourable therapy for both its well-tolerated profile and clinical benefit in patients with HCC adjacent to the major vessels. Moreover, this treatment was equally effective as wide-margin hepatectomy. However, the patient population is too small to make a definitive conclusion; prospective studies with larger patient cohorts are needed to confirm these results.Part Ⅱ:Contribution of Phenotypic Heterogeneity of pT1-2N0 Invasive Breast Cancer to Individual Prognosis:A Large Cohort Study with Cluster AnalysisObjective:To find phenotypic subgroups of patients with pTl-2N0 invasive breast cancer by means of cluster analysis and estimate the prognosis and clinicopathological features of a specific subgroup.Material and methods:From 1999 to 2013,4979 patients with pT1-2N0 invasive breast cancer were recruited for hierarchical clustering analysis. Age (<40,40-69,70+ years), diameter of primary tumour, histological subtype, grade of differentiation, micro vascular invasion, oestrogen receptor (ER), progesterone receptor (PR) and human epidermal growth factor receptor 2 (HER-2) were chosen for distance metric between patients. Hierarchical linkage method was Ward’s method. Cophenetic correlation coefficient (CPCC) and Spearman correlation coefficient is used to validate clustering structures.Results:CPCC was 0.603; Spearman correlation coefficient was 0.617 (p< 0.0001), which indicated a good fit of hierarchy to the data. A twelve-cluster model seemed to best illustrate our patient cohort. Patients in Cluster 5,9 and 12 had best prognosis and was characterized by age≥40 years, smaller primary tumor, lower grade, positive ER and PR, and mainly negative HER-2. Patients in Cluster 1 and 11 had worst prognosis, Cluster 1 was characterized by larger tumor, higher grade and negative ER and PR, while Cluster 11 was characterized by positive microvascular invasion. Patients in other 7 clusters had moderate prognosis, and patients in each cluster had distinctive clinicopathological features and recurrent patterns.Conclusion:This study identified distinctive clinicopathologic phenotypes in a large cohort of patients with pT1-2N0 breast cancer through hierarchical clustering and revealed the different prognosis. Then, this integrative model would help physicians make informed decisions regarding the personalized adjuvant therapy.Part III:Application of Stacked Denoising Autoencoder-Based Survival Analysis on pTis-3N0-3M0 Breast Cancer DatasetsObjective:To apply deep learning to a non-linear prediction model of recurrence for patients with breast cancer.Methods:From 1999 to 2013,5670 patients with pTis-3N0-3 breast cancer were recruited for a prediction model. The dataset was randomly divided into training subset and validating subset by a ratio of 4:1. The deep learning model is constructed using training subset and tested on the validating subset to obtain its error rate. Independent variables include patient age, status of neoadjuvant chemotherapy, histological subtype, diameter of primary tumour, axillary lymph node status, histological grade, microvascular invasion, ER, PR and HER-2. Adaptive learning with 15 epochs was used for pre-training and 1000 epochs at most used for fine tunning. Area under the receiver operating characteristic (ROC) curve (AUC) was measured to estimate accuracy of model.Results:Among 5670 patients, recurrences were observed in 1308 cases. Four hundred and eighty-seven of the 1308 encountered locoregional recurrence during 5-year follow-up. There were significant difference between cases with and without locoregional recurrence in patient age, diameter of primary tumour, axillary lymph node status, metastatic ratio of lymph node, status of neoadjuvant chemotherapy, ER, PR and HER status. There were significant difference between cases with and without any recurrence in patient age, diameter of primary tumour, axillary lymph node status, metastatic ratio of lymph node, status of neoadjuvant chemotherapy, and PR status. Due to lower level of hormone receptor expression, patients with recurrence had a lower probability of receiving adjuvant endocrine therapy. Moreover, patients experiencing recurrences had larger primary tumour, more lymph nodes metastasis and higher metastatic ratio of lymph node, hence, more of them received adjuvant radiotherapy. Finally, area under ROC curve of prediction models for 5-year disease-free survival and logoregional recurrence-free survival were 0.73 and 0.72, respectively.Conclusion:We trained a prognostic model with stacked denoising autoencoders network to individually predict 5-year disease-free survival and logoregional recurrence-free survival of patients with breast cancer, what’s more, the work laid a good foundation for decision support system of personalized treatments.