Abstract Introduction: High dimensional imaging approaches such as imaging mass cytometry (IMC) are becoming widely used in cancer research. Such methods allow simultaneous observation of many cell types and their functional states and can provide valuable spatial information on cancer disease states when applied to clinical tissue samples. For example, in-situ immune and tumor cell interactions can be interrogated in their spatial context within the tumor microenvironment (TME). Analysis methods for the resultant complex data are not well formalized, and bespoke methods are usually required to fully capitalize on the underlying richness of information made available by IMC. Deep learning [DL] approaches, while highly accurate for other imaging modalities, have been slow to be adopted in IMC, as public resources for deep learning tasks in IMC are not abundant. Methods: We developed multiple DL and ML-based analysis pipelines for the following tasks in IMC data processing: [1] nucleus and necrotic tissue segmentation, [2] quantitative nuclear and cellular morphometry, [3] identification of cell type-specific niches. We applied these protocols to images and derived single cell spatial data from the TRACERx IMC cohort (n=81 non-small cell lung cancer patients, 561 images). Results: [1] We created a 120 image, 46,000+ labelled nucleus segmentation dataset for IMC data with representative images from lung adenocarcinoma, squamous cell carcinoma and other tissues. We achieved state-of-the-art performance in nuclear instance segmentation using a custom U-net++ neural network architecture trained using this dataset, which we benchmarked against traditional image processing methods, as well as publicly available deep learning architectures. Subsequently, we exploited transfer learning to retrain this model on a restricted dataset of labelled necrotic domains, which produced predictions in good agreement with independent pathologist assessment. [2] We developed an IMC morphometry pipeline utilizing ML-informed partitions of nuclear and cellular shape descriptors through which we performed cell-type specific morphometric characterization of all mapped cells in the non-small cell lung cancer TME, and which enabled a comparative analysis of the morphometries of each distinct cellular phenotype. [3] We established a high throughput density-based spatial clustering pipeline capable of identifying locally enriched niches of a given cell type of interest, as well as probing the composition and phenotypes of other cells within these niches. Conclusions: These approaches enhanced the quality as well as the breadth of spatial information derivable from TRACERx IMC data. Applying such tools to other clinical and pre-clinical datasets can improve our understanding of the spatial organization of cells both in non-small cell lung cancer and other cancer types. Citation Format: Alastair Magness, Katey Enfield, Mihaela Angelova, Emma Colliver, Emer Daly, Kristiana Grigoriadis, Claudia Lee, Oriol Pich, Philip Hobson, Dina Levi, Takahiro Karasaki, David Moore, Julian Downward, Erik Sahai, Mariam Jamal-Hanjani, Charles Swanton, TRACERx Consortium. Machine learning-enhanced image and spatial analytic pipelines for imaging mass cytometry applied to the TRACERx non-small cell lung cancer study [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2022; 2022 Apr 8-13. Philadelphia (PA): AACR; Cancer Res 2022;82(12_Suppl):Abstract nr 1926.
Read full abstract