Quantifying Tumor Infiltrating Lymphocytes in Ovarian Cancer TMAS.
Andrew Janowczyk, Sharat Chandran, Michael Feldman, Anant Madabhushi. IIT Bombay, Mumbai, India; UPenn, Philadelphia; NJ Rutgers, New Brunswick
Background: Research has shown that the presence of lymphocytes within tumor cells (tumor infiltrating lymphocytes (TILs)), is reflective of outcome for Ovarian Cancer (OCa). Hence, is it important to distinguish between TILs and lymphocytes that reside in the stroma (non-TILs). Manually counting TILs is tedious and time consuming, thus there is a clear need for automation. Standard computer vision approaches are not only computationally expensive, but fail to account for subtle differences in local structural morphology (TILs are densely surrounded by tumor cells, non-TILs are not). We present a novel local morphologic scale (LMS) descriptor which quantitatively captures local structure. Morphologic features derived from LMS can be used to train a supervised classifier to distinguish between stroma and tumor regions, and thus TILs and non-TILs.
Design: The algorithm is as follows: Identify all lymphocytes from a stained OCa TMA using Hierarchical Normalized Cuts (HNCut), a minimally interactive object detection scheme. Next we generate LMS signatures (Figures 1 (b), (d)) for regions surrounding each lymphocyte. This signature is determined by sampling the path of “particles”, which are adjusted based on a simple physics model, as they leave the region. Features (such as length of LMS path) at each location can then be extracted and used to train a classifier to distinguish between TILs and non-TILs.
Results: 16 TMA cylinders across 6 different OCa studies were used to evaluate the LMS algorithm against a pathologist's annotation of what constituted a TIL. The LMS classifier was employed to identify TILs within a total of 8000 200X200 image patches. A Probabilistic Boosting Tree was trained using LMS features to distinguish between TILs and non-TILs using 80% of the data, the remaining 20% being used for evaluation. Average ROC AUC was .803.
Conclusions: We presented a novel, local, morphologic image descriptor (LMS) that was successfully applied to rapidly and automatically discriminate TILs and non-TILs on OCa TMAs. In future work we intend to apply LMS to identifying TILs in the context of other diseases (e.g. Her2+ breast cancers).
Tuesday, March 1, 2011 1:00 PM
Poster Session IV # 190, Tuesday Afternoon