Abstract LB-368: Applications of automated image analysis, machine learning and spatial statistics for the improvement of stage II colorectal cancer prognosis

Ines P. Nearchou, Daniel A. Soutar, Kate Lillard, Hideki Ueno, Oggie Arandelovic, David James Harrison, Peter David Caie

Research output: Contribution to conferenceAbstractpeer-review


Background and Objectives: The tumor microenvironment (TME) plays an important role on tumor progression and patient survival outcome. The TME varies significantly amongst patients as well as within individual tumors. Although a number of studies have reported the prognostic significance of the various TME components, only a very small number of those address the issue of intra-tumor heterogeneity. In this study, we evaluate the densities and interactions of tumor infiltrating lymphocytes, macrophages and tumor buds (TBs) in order to create a more personalized prognosis for patients with stage II colorectal cancer (CRC). This was achieved through the use of multiplexed immunofluorescence, automated image analysis and machine learning approaches. In addition, we developed an objective methodology for studying the intra-tumor heterogeneity and assess its impact on patient survival outcome. Methods: Multiplexed immunofluorescence and automated image analysis using HALO® software were applied for the quantification of CD3+, CD8+ T cells, CD68+, CD163+ macrophages and TBs, across 2 sequential whole slide images (WSI). This was performed on 230 stage II CRC patient samples from Scotland and Japan. Density and spatial relationships between the cellular subpopulations were averaged across the WSI to form input for a prognostic model. To evaluate the intra-patient heterogeneity a further analysis method was developed which divided the WSI into grids with a fixed tile area of 0.785mm2. Tiles with significantly small or large numbers of the feature of interest were considered hot or coldspots respectively. The number of each objects' hot or coldspots within each patient were then calculated. Two machine learning algorithms were employed for the analysis of the data from each analysis method, which lead to the development of two new prognostic risk models. Results: The first combinatorial prognostic model, utilizing the averaged data, consisted of lymphocyte infiltration, the number of lymphocytes within 50µm of TBs and CD68+ /CD163+ macrophage cell ratio. This model was shown to identify a subpopulation of patients who exhibit 100% survival over a 5-year follow-up period. This finding was confirmed in an independent and international validation cohort. The second prognostic model using the results from the spatial heatmap analysis, included the number of TB hotspots as well as the number of hotspots for the proximity of lymphocytes to TBs. This model was shown to be of high prognostic significance. Conclusion: This work demonstrates how by applying digital pathology and machine learning approaches it is possible to identify stage II CRC patients for whom surgical resection alone may be curative. Furthermore, we report a new methodology to evaluate the intra-tumor heterogeneity which was found to improve stage II CRC patient stratification when compared to the current clinical gold standards.
Original languageEnglish
Publication statusPublished - Aug 2020


Dive into the research topics of 'Abstract LB-368: Applications of automated image analysis, machine learning and spatial statistics for the improvement of stage II colorectal cancer prognosis'. Together they form a unique fingerprint.

Cite this