Immune profiling of the colorectal cancer microenvironment for precision prognostics

Abstract

The tumour-node-metastasis staging system is the gold standard for colorectal cancer (CRC) patient stratification into prognostic subgroups and for treatment decision making. However, same-stage patients may have a different clinical outcome. Several features within the CRC microenvironment have previously been shown to have high prognostic significance. However, they are traditionally reported independently. This thesis aimed to assess the prognostic significance of tumour budding, lymphocytic and macrophage infiltration and desmoplastic reaction (DR), as well as their spatial inter-relationships across whole slide tissue sections through the application of automated image analysis and machine learning (ML) methodologies.

Firstly, three image analysis algorithms were developed to automatically quantify tumour buds (TBs), lymphocytic and macrophage density on immunofluorescence whole slide images of patients from Scotland (n = 170) and Japan (n = 62). Survival analysis revealed that all were highly significant prognostic factors. The prognostic value of their spatial inter-relationships was also assessed, and results revealed that a low number of lymphocytes within 50μm of TBs was associated with reduced disease-specific survival.

Secondly, ML algorithms were used for the development of two novel prognostic risk models: the tumour bud-immuno spatial index (TBISI) and the spatial immuno-oncology index (SIOI). The TBISI, which integrated TBs, lymphocytic infiltration and their spatial inter-relationship, was shown to be more accurate in patient stratification than any of the features in isolation or pT stage. The SIOI, which derived from integrating lymphocytic infiltration, the spatial association of lymphocytes and TBs and CD68⁺/CD163⁺ macrophage ratio, identified a subpopulation of patients who exhibit 100% survival over a 11.8-year follow-up period. All the aforementioned findings were confirmed in independent validation cohorts.

Finally, a novel ML algorithm for the automatic and standardised classification of DR on stage II and III CRC patients (n = 528) was developed. The algorithm was shown to successfully classify DR on unseen whole slide images. Survival analysis results revealed that the ML classifier had a higher prognostic significance than manually assessed DR.

In conclusion this thesis demonstrates how the use of automated image analysis can be successfully used for the standardised and reproducible assessment of various features and their inter-relationships within the complex tumour microenvironment. Additionally, it demonstrates how through ML approaches it is possible to develop novel combinatorial prognostic models which improve the prognostic accuracy in CRC.