Target
The type and function of a cell is related to protein distribution in this cell. Hence, the target of the project#2 is to train a neural network which can segment cells in an image of proteins under the microscope with other images as references.
Data
Inputs of a sample are four microscope images corresponding to same cells. Four images are the protein of interest (green), nucleus (blue), microtubules (red), endoplasmic reticulum (yellow) respectively. Due to the target of this project which finds the mapping between protein distribution and kinds of cells, green images should be employed to predict labels while other images are considered as references. There are 19 different labels indicating types of cells (The labels 0-17 are different cells, the label 18 is for negative and unspecific cells)
As this is a weakly supervised learning task, although labels of a sample in the training set is image-level label which indicates what are kinds of cells in the sample, the target of the test set is to segment the cells in the images (lists of instance binary segmentation masks) and predict the labels of those segmented cells. Here is an example:
(c) microtubules(red) (d) endoplasmic reticulum(yellow)
Fig 1. A sample of input images with the train label 16 | 2
(a) the mask filtering out cell 16 (b) the mask filtering out cell 2
Fig 2. Desired binary segmentation masks when test
Evaluation
For each image in the test set, you must predict a list of instance binary segmentation masks and their associated detection score (Confidence). To save the storage, each mask is required to be encoded to a string via the code in the Kaggle website-> Overview->Evaluation.
