(Wednesday, 7/17/2019) 10:30 AM - 11:00 AM
Room: Exhibit Hall | Forum 5
Purpose: To automatically detect and segment lung tumors in simulation CT scans for lung radiotherapy treatment planning.
Methods: A 2D convolutional neural network (CNN) was trained to automatically detect and segment lung tumors in CT scans. CT scans from 2650 patients were selected to train (2154 patient scans) and test (496 patient scans) the model. Clinically-approved GTV contours were used as ground truth. The model was trained on a slice-by-slice basis with a predefined lung mask. 2D predictions in each slice were combined to create 3D predictions for each patient. Each 3D connected region was considered a tumor candidate. CNN predictions could generate multiple tumor candidates for each patient. Then a manual review was performed to filter out false positives among the tumor candidates. The selected tumor candidate was used to assess the detection performance of the CNN model. To assess the segmentation accuracy, the Dice and mean surface distance (MSD) were calculated between the selected tumor candidate and the ground truth.
Results: The range of tumor volumes in the test set was from 0.37cc to 46.76cc (average=8.24cc; SD=8.20cc). The average number of false positives per test case was 17.57. Under physician review, 95%, 79%, and 57% of tumors were correctly detected by the CNN model for tumor volumes greater than 10cc (n=132), between 5cc and 10cc (n=134), and less than 5cc (n=230), respectively. For the test cases that were correctly detected, the average MSD was 2.49±4.29mm, 1.58±1.00mm and 1.13±0.56mm and the average Dice was 0.64±0.18, 0.57±0.22 and 0.53±0.25 for tumor volumes greater than 10cc, between 5cc and 10cc, and less than 5cc, respectively.
Conclusion: Our CNN model was able to identify and segment the large majority of lung tumors; further work is ongoing to automatically reduce false positive predictions without the need of user intervention.
Keywords
Segmentation, CAD, Image Analysis
Taxonomy
Contact Email
