Ductal carcinoma in situ (DCIS) is a non-obligatory precursor of invasive ductal carcinoma (IDC). It is the most common mammographically detected breast cancer, however, predicting DCIS progression to IDC remains a major clinical challenge. A recent study has categorised DCIS evolution to IDC into four models, highlighting its heterogeneity. The evolutionary potential of individual DCIS ductules/ducts may dramatically differ, determined by not only their genetic mutations but also microenvironmental selective pressure. However, given the complex spatial ductule structure, ecological dynamics between individual DCIS ducts and their surrounding microenvironment are difficult to measure by eye. These ultimately limits our ability to study the influence of the microenvironment on tumour evolution and progression.

Our primary aims were: (1) to develop and validate a computational pipeline that accurately detects and segments individual DCIS ducts; (2) to characterise the immune microecology for each DCIS duct using spatial statistics on H&E and IHC for TILs; (3) to test the difference in DCIS microecology between samples with pure DCIS and DCIS samples derived from IDC patients (adjacent DCIS, as a surrogate for poor prognosis DCIS).

UNMaSk pipeline consists of processing pipelines for segmentation of ductal carcinoma in situ (DCIS) regions from whole slide images. The primary steps are broken down into four modules and are as given below.

• HE_tissue_segmentation

• DCIS segmentation

• HE_cell_detection

• HE_cell_classification

Modular arrangement help in navigation to each module and supports both tile level and tissue level processing. Each of these pipelines are organised inside individual directory and you will be able to find detailed explanation in the respective sub-directories. Wherever possible docker images and command line instructions are specified to make it user friendly for off the shelf users.

a. Images used for training https://github.com/pathdata/UNMaSk/tree/master/DCIS/TrainData (Left)

b. Ground truth images https://github.com/pathdata/UNMaSk/tree/master/DCIS/TrainData/mask (Middle)

c. Overlay of groundtruth on the training image https://github.com/pathdata/UNMaSk/tree/master/DCIS/TrainData/overlay (Right)

Priya Lakshmi Narayanan et al., (2021) Unmasking the immune microecology of ductal carcinoma in situ with deep learning. https://doi.org/10.1038/s41523-020-00205-5

All training data of carcinoma in situ regions that were annotated as a part of the project is made available in this github repository. Training data tiles were anonymised from raw HE image tiles. Request for data access for the Duke samples can be submitted to E.S.H and Y.Y

1. This project is a work in progress and contact the corresponding authors for any queries. Docker images are packaged in each of these modules are made available and few example test images are provided for users to start their experiment. All the modules are independant of each other so each module can be iteratively tested. Usage of tiled images will help to test the pipeline without performing tissue segmentation.