Abstract:
Computed Tomography (CT) is the most widely used imaging procedure for locating
and diagnosing kidney tumors. The standard treatment for kidney tumors is surgical
removal. It is important to accurately segment the kidney and its tumor for effective
surgical planning. The manual segmentation of kidney tumors is time-consuming and
subject to variability between different radiologists. Therefore, automatic semantic
segmentation of kidney tumors using deep learning networks has become increasingly
popular in the past few years. Automatic segmentation of kidney tumors is a very
challenging task due to their morphological heterogenicity. This work provides the
application of 3D UNet and 3D SegResNet on KiTS19 challenge data for accurate
segmentation of kidney and kidney tumors. An ensembling operation was added in the
end to average the predictions of all models. The proposed method is based on the
MONAI framework and focuses more on training procedure rather than complex
architectural modifications. The models were trained using KiTS19 training set of 210
cases for which ground truth labels were available. The training data was divided into
190:20, for training and validation respectively. We evaluated the performance of our
network on KiTS19 official test set and obtained mean dice of 0.8964, 0.9724 kidney
dice, and 0.8204. Our approach outperforms many submissions in terms of kidney
segmentation and gives promising results for tumor segmentation. We also used a local
test set of 90 cases from KiTS21 challenge to check how well our method adopts to a
new dataset. It scored a mean dice of 0.9160, kidney dice of 0.9771, and 0.8550 tumor
dice. The obtained results on KiTS19 official test set and local test set show that our
approach is effective and can be used for organ segmentation.
