Segmenting the spleen in CT-images using deep learning algorithms. Automatically reporting splenic size and change in size during follow-up. Evaluation of its usefulness in the clinical setting.
PURPOSE:
To develop a fully automatic deep learning method for 3D segmentation of the spleen on computed tomography (CT) scans and to compare the automatically measured spleen volume with the standard splenic index approximation formula that requires three 2D manual measurements by a radiologist. The usefulness of this measurement in aiding an experienced radiologist to judge splenic volume change in the clinical setting was also evaluated.
METHODS:
2150 CT thorax-abdomen scans were collected from the Oncology department of the Radboud University Medical Center. 1100 of these were used to develop the algorithm, 50 as a test set for the segmentation task, and 1000 to validate the clinical usefulness of the algorithm. A state-of-the-art 3-D convolutional neural network was used for segmentation.
RESULTS:
The deep learning network reached a dice score of 0.9623 on the test set of 50 scans, while a second observer obtained a comparable dice score of 0.9638. In the observer experiment, the radiologist scored the quality of the predicted segmentations as excellent in 87% of the cases, 7% good (up to 5mm error), 3.5% bad, and 2.5% segmentation failure. After a first visual classification of volume change, the radiologist changed his classification when seeing the output of the network in 15% of the cases.
CONCLUSIONS:
State-of-the-art segmentation methods based on deep learning can accurately segment the spleen in CT scans. This may be integrated into clinical workstations to detect abnormal splenic volumes and splenic volume changes over time.
