Task 3: mitotic cell recognition

Description

Cells undergoing mitotic phase will express different amount of antibodies which would be a useful sign for narrowing down the possible patient pattern. Before determining the mitotic cells, the cells are first detected. As mentioned, the scientists need to ensure that there are at least one or two mitotic cells on each patient’s specimen.

Due to their much lower number of occurrences, we name the second task as the mitotic cell detection. In this setting, we assume the interphase cells (i.e. the non-mitotic cells) as the “background” cells.

Recently it was found that mitotic cell detection can addressed using the secondary counterstain. Again, the introduction of secondary counterstain may not be economical. Thus, in this work we only primarily aim on the problem of mitotic cell detection using the primary counterstain.

 

Dataset

The images were acquired between 2014-2015 at the Sullivan Nicolaides Pathology Laboratory, Australia . It was collected from 253 anonymized patients. Each patient serum was prepared as per industry protocol with 1:80 titer dilution and the specimen was photographed using a monochromatic camera fitted on a microscope. From 253 patient specimens, approximately 23,000 interphase cells and 1,000 mitotic cells were extracted.

Each image contains the following information:

• The groundtruth label: mitotic (+1) and interphase (-1)

 

Evaluation

Due to low number of mitotic cells compared to the whole cell distribution, the mitotic cell recognition can be considered as a detection problem. Here, the interphase cells are considered as a background and the mitotic cells are considered as the cell of interest. More specifically, we define the evaluation metric using the standard object recognition/detection problem. In particular, the evaluation will utilize the detection confidence score of each image provided by the submission. The confidence score indicates the confidence of an cell image belongs to the mitotic cell. The evaluation will be determined using two means: (1) the precision-recall curve and (2) the Average Precision (AP) which is calculated based on the precision-recall curve. The team that achieves the best on both metric will be declared as the winner. The proclamation of the winner will be made during the const session at the ICPR 2016.