Ghannam, Suzan
(2025)
Deciphering the morphological features underlying breast cancer behaviour using computational pathology and Artificial intelligence (with focus on the role of the tumour microenvironment).
PhD thesis, University of Nottingham.
Abstract
Background:
Breast cancer (BC) is a heterogeneous disease with variable presentations, morphologies, and behaviours. The tumour microenvironment (TME) plays a key role in tumour progression. Molecular and phenotypic alterations in the neoplastic cells and their behaviour usually attracts more attention. However, there is evidence that the stroma of BC plays an important role in tumour behaviour. The basement membrane (BM), which comprises a physical barrier that separates the proliferating epithelial cells from the surrounding stroma is used to distinguish malignant in situ and invasive lesions. However, exceptions exist. The stroma of BC also varies widely with a variable degree of fibrosis, collagenisation, and elastosis in addition to other microenvironment components such as immune cells. The stromal cells are relatively genetically stable compared to the tumour cells, therefore they may be targets for therapeutic agents for cancer. The revolution of using digitalised whole slide images (WSI) in pathology developed a way of investigating BC cases. In addition, using artificial intelligence (AI) enables the integration of knowledge beyond human limits. However, this is a recently introduced technology, so further assessment of its uses and applications in the clinical setting is of paramount importance. Therefore, this study integrates digital pathology (DP) and AI to decipher the morphological features of TME in BC and identify novel biomarkers for improved BC prognosis.
Patient and Methods:
This study included multiple large cohorts of primary BCs and ductal carcinoma in situ (DCIS) in either tissue microarray (TMA) or full-face sections. The Nottingham cohort and the online cohort the Cancer Genome Atlas (TCGA) BC dataset. The detailed clinicopathological and outcome data were collected for the Nottingham cohort. For the DP part of the study, the haematoxylin and eosin (H&E)-stained slides were histologically reviewed and scanned into WSIs, and the same sections were picrosirius (PSR) stained to be examined under polarised microscopy or kept unstained for differential interference contrast microscopy (DIC). Image analysis assessment of the geometric characterisation of collagen was performed using AI applications.
For the immunohistochemical (IHC) part of the study, TMAs were constructed, and full-face sections were also utilised. Mining for stromal-related genes and collagens-related genes as novel biomarkers that may have potential prognostic and predictive values in BC was performed. This included hypoxia-inducible factors alpha 1 (HIF-1α), Lysyl oxidase (LOX), collagen XV and XIX.
Differential gene expression (DGE) analysis was performed to identify a set of genes associated with stroma in the main molecular BC subtypes and their prognostic roles in each molecular subtype. Stromal features and types were scored in 822 cases and the association to outcome was evaluated.
Results:
This study aimed to decipher the morphological features of TME in BC. Therefore, we started with BM as a part of TME and hypothesised that differentiation of a native BM from a reactive BM surrounding an invasive lesion is clinically important for early diagnosis. The study included 150 cases divided into six groups, each containing 25 patients per group. The six groups included normal breast tissue as a control group, DCIS, encapsulated papillary carcinoma (EPC), invasive carcinoma (INV), special types of invasive carcinoma (S.INV) and metastatic lymph node (LN). Full-face sections were stained with PSR stain and examined using polarized microscopy. Image analysis software was used to assess collagen fibre parameters. DCIS’ BMs resembled normal tissue BMs except the collagen fibres were higher in density, straighter, disorganised and less aligned. However, EPC, invasive and special invasive groups had a higher density of collagen with thicker fibres, and more collagen I content which were wider, shorter, straighter, less aligned and disorganised. The EPC group resembled invasive groups regarding alignment but had longer fibres than special BC but less wide fibres, and less collagen I (thick fibre) content, than invasive BC.
Therefore, the EPC group was further analysed to assess the differences between the inner and outer parts of the thickened capsule and to compare the capsule parts with the other groups with an additional encapsulated papillary thyroid carcinoma (EPTC) control group. The collagen fibre parameters of EPC capsules differed from the BMs of normal breast tissue and DCIS, and EPC was surrounded with BM-like material resembling those surrounding some invasive tumour types. This provides further evidence that EPC is a reactive process rather than a thickened native BM. In addition, most papillary tumours in different organs, such as EPTC which are surrounded with BM-like material showed indolent behaviour with better prognosis.
To study collagen fibre parameters of the stroma, a TMA was generated from 200 cases, comprising ductal carcinoma in situ (DCIS; n=100) and invasive tumours (n=100), with an extra 50 (25 invasive BC and 25 DCIS) cases for validation was unstained for examination using DIC microscopy. The collagen fibres had higher density, and were thinner, straighter more disorganized and less aligned in the invasive BC compared to DCIS. A model considering these features was developed that could distinguish between DCIS and invasive tumours with 94% accuracy. There were strong correlations between fibre characteristics and clinicopathological parameters in both groups. A statistically significant association between fibre characteristics and patients’ outcomes was observed in the invasive group but not in DCIS.
Stromal-related genes and collagen-related genes were selected for immunohistochemistry assessment to explore novel biomarkers. The assessment of the selected biomarker panel revealed a significant association between their protein expression and patient outcome. For the assessment of collagen XV and XIX, 100 cases of full-face sections were used which included both normal and neoplastic breast tissue. The cases were divided into four groups, each containing 25 patients. The four groups included normal breast tissue, DCIS, EPC, and invasive carcinoma. Collagen XV and XIX were significantly involved in the structural and biological changes associated with BC and may act as potential biomarkers. Collagen XIX could be used as a diagnostic tool to differentiate between invasive and non-invasive lesions in challenging cases. For the assessment of HIF-1α and LOX protein expression, TMA sections of a large well-characterised cohort of BC (n=876) were used. HIF-1α and LOX could be new biomarkers for BC prognosis and therapy response.HIF-1α levels in BC may be used to prospectively stratify patients who received neoadjuvant therapy. LOX could be a good therapeutic target specifically for metastasis prevention.
The mechanisms of stroma formation and composition in different molecular subtypes, which could explain the different prognostic values, were evaluated. Two large well-characterized BC cohorts were used, an in-house BC cohort (n=822) and the public domain dataset TCGA, n=978) as a validation cohort and for differential gene expression (DGE) analysis. DGE was performed to identify a set of genes associated with high stroma tumour ratio (STR) in the three main molecular subtypes. In each subtype, stromal assessment was carried out and tumours. were assigned to two groups: high and low STR, and further correlations with tumour characteristics and patient outcomes were investigated. The contribution of tumour-infiltrating lymphocytes (TILs) to the stroma was also studied. High STR was associated with favourable patient outcomes in the whole cohort and the luminal subtype, whereas high STR showed an association with poor outcomes in triple-negative (TNBC). No association with outcome was found in the HER2-enriched BC. DGE analysis identified various pathways in luminal and TNBC subtypes, with immune upregulation and hypoxia pathways enriched in TNBC, and pathways related to fibrosis and stromal remodelling enriched in the luminal group instead. Low STR accompanied by high TILs was shown to carry the most favourable prognosis in TNBC. In line with the DGE results, TILs played a major prognostic role in the stroma of TNBC, but not in the luminal or HER2-enriched subtypes. The stroma type could be a predictor factor for predicting outcomes. Desmoplastic sclerotic stroma is the most common type and showed a worse prognosis compared to other stromal types in all histological types except lobular carcinoma.
Conclusion
This comprehensive study emphasises the importance of considering all morphological features of TME in tumour behaviour and prognosis. Collagen parameters of BM and stroma could be used as a diagnostic and prognostic tool to differentiate between invasive and non-invasive lesions. Most papillary tumours in different organs which are surrounded with BM-like material showed indolent behaviour with better prognosis, therefore mechanisms which explain these phenomena should be further investigated. Stromal and collagen-related genes could be used as prognostic biomarkers for BC. The underlying molecular mechanisms and composition of the stroma in BC are variable in the molecular subtypes which explain the differences in its prognostic significance in each molecular subtype. Evaluation of tumour stroma types and features may be easily performed in routine clinic practice as a guide in stratifying patient risk.
Actions (Archive Staff Only)
 |
Edit View |
Tools
Tools
