Breast composition: relationship with tumour characteristics and treatment outcomes

Abstract

Abstract Aims: High mammographic density (MD) increases the risk of breast cancer and plays a role in breast cancer progression. However, the relationship between MD and tumour characteristics and treatment outcome is poorly understood. The aims of the study include: 1) examining the relationship between mammographic density phenotypes and breast cancer (BC) characteristics and tumour location; 2) exploring the abilities of global textural features extracted from the ipsilateral breast mammograms to predict BC characteristics; 3) evaluating the prognostic utility of baseline MD in BC patients; 4) investigating changes in MD over time following BC treatments, and exploring factors affecting these changes. Methods: The work presented in this thesis was conducted in four stages. In the first stage, 297 women with BC were included, and MD was qualitatively using the breast Imaging-Reporting and Data System and quantitatively using the Laboratory for Individualized Breast Radiodensity Assessment-LIBRA measured from contralateral breast mammograms. Using radiology report descriptions, surgical scars, and visible metallic markers placed on mammograms, approximate tumour locations were determined. Binary logistic regression analysis was employed to evaluate the association between MD and tumour characteristics. In Stage 2, segmentation was performed prior to feature extraction. Using MATLAB R2018a software, a total of 33 global radiomic features were extracted from the ipsilateral breast mammograms of 184 females diagnosed with BC. Univariate logistic regression analysis was performed to select radiomic features and a Receiver Operating Characteristics (ROC) Curve analysis was conducted to assess the discriminatory power of the global radiomic features from the ipsilateral breast mammograms in predicting BC characteristics (oestrogen hormone receptor, progesterone 4 hormone receptor, human epidermal growth factor receptor 2, tumour invasiveness, lymph node status, Nottingham histological grade, and tumour size). In Stage 3, two different fully automated MD assessment methods (AutoDensity and LIBRA) were used to assess MD at BC diagnosis for 224 women with BC, and their output was used to categorise patients into two groups: the low MD (PD<20%) and high MD (PD≥20%) group. Kaplan-Meier analysis and the Cox-proportional hazard models were then employed to investigate the prognostic utility of baseline MD. In Stage 4, MD of 226 BC affected women was quantitatively evaluated by the LIBRA software before (at BC diagnosis) and after BC treatment initiation. A maximum of six follow-up mammograms were selected to monitor MD changes following BC treatment (mean: 71.29 months, range: 69-73 months). The Wilcoxon ranked signed test was used to examine the differences in MD changes and the median test was used to examine the factors influencing these changes within one year of treatment initiation. All mammograms used in this thesis were digital mammograms. Results: In Stage 1, no MD phenotypes (numerical and categorical variables) showed statistically significant association with BC characteristics. All analyses demonstrated P-values equal to or greater than 0.05. BC was more likely to develop in dense tissues with distinct BC features including human epidermal growth receptor 2 (p=0.05) and carcinoma in situ (p=0.01). In Stage 2, progesterone hormone receptor status was weakly discriminated by two histogrambased features (mean, and 70th percentile) (AUC range: 0.650-0.652, p ≤ 0.003), and tumour size by one histogram-based feature (30th percentile) (AUC: 0.627, p = 0.007). Similarly, the grey level run length matrix (GLRLM)-based feature (grey level non-uniformity) poorly predicted lymph node status (AUC: 0.68, p = 0.007), and the fractal dimension showed low predictive power for tumour size (AUC: 0.