The data we use is anonymized MRIs of patients treated with MRI guided radio therapy provided by the UW Madison Carbone Cancer Center. Specifically, the dataset contains 85 cases with 38496 scan slices of organs represented in 16-bit grayscale PNG format, and the annotations are provided in a csv format with the segmented areas represented as RLE-encoded masks. Sample csv annotations are shown in Table 1. An empty segmentation entry represents no mask presented for the class in the MRI scan slice.
We use the Dice score to evaluate the mask predicted by our model against the true mask. As described above, the Dice score of a predicted mask is defined by 2|Mpred ∩ Mtrue| |Mpred| + |Mtrue| where Mpred is the set of pixels masked in (i.e. predicted to class ”1”) by the model, Mtrue is the set of pixels masked in by ground truth mask, and the absolute value means the number of pixels.
In this project we are segmenting organs cells in images. The training annotations are provided as RLE-encoded masks, and the images are in 16-bit grayscale PNG format.
Each case in this project is represented by multiple sets of scan slices (each set is identified by the day the scan took place). Some cases are split by time (early days are in train, later days are in test) while some cases are split by case - the entirety of the case is in train or test. The goal of this project is to be able to generalize to both partially and wholly unseen cases.
Note that, in this case, the test set is entirely unseen. It is roughly 50 cases, with a varying number of days and slices, as seen in the training set.