Click here to


Are you sure ?

Yes, do it No, cancel

Convolutional Neural Networks for Fully Automatic Segmentation of Lung Tumors in CT Images

C Owens1,2*, C Peterson2,3 , C Tang4 , E Koay4 , W Yu5, J Li4 , M Salehpour1 , D Fuentes6 , L Court1,2,6 , J Yang1,2 , (1) Department of Radiation Physics, The University of Texas MD Anderson Cancer Center, Houston, TX, (2) The University of Texas Graduate School of Biomedical Sciences at Houston, Houston, TX, (3) Department of Biostatistics, The University of Texas MD Anderson Cancer Center, Houston, TX, (4) Department of Radiation Oncology, The University of Texas M.D. Anderson Cancer Center, Houston, TX, (5) Department of Radiation Oncology, Shanghai Chest Hospital, Shanghai Jiao Tong University, Shanghai, China, (6) Department of Imaging Physics, The University of Texas M.D. Anderson Cancer Center, Houston, TX


(Tuesday, 7/31/2018) 9:30 AM - 10:00 AM

Room: Exhibit Hall | Forum 5

Purpose: To automatically segment lung tumors in CT images using deep convolutional neural networks (CNN).

Methods: Nine CT scans of lung cancer patients were selected to train a 2D U-Net CNN for lung tumor segmentation. Three radiation oncologists manually delineated each lung tumor twice. A group-consensus contour was generated using majority voting from the manual contours and used as ground-truth for CNN model training. Of the 9 CT scans, 79 slices out of 1258 contained tumor labels. Additionally, a lung mask was automatically generated for each patient and used for model training as well. To mitigate the effects of high class-imbalance, our CNN model used a weighted cross-entropy cost function. The tumor prediction from the model was further processed automatically by removing dense pulmonary vessels to produce the final segmentation. We used leave-one-out cross-validation to assess the segmentation accuracy. The Dice similarity coefficient (DSC), true positive rate (TPR) and false positive rate (FPR) were generated between the auto-segmented CNN contour and the group-consensus contour for each patient.

Results: The median DSC of the 9 patients was 0.61±0.20 and 0.74±0.36 before and after post-processing was applied to the model prediction, respectively. The mean TPR was 0.81±0.11 and 0.63±0.35, and the mean FPR was 0.0016±0.0012 and 0.0007±0.0008 before and after post-processing, respectively. Of the 9 patients, one had a very small tumor volume and one contained a large amount of contrast-enhanced vessels. The model prediction resulted to low DSCs (0.19 and 0.45 respectively) and tumors were completely wiped out after post-processing. After removing these two cases, the median DSC increased to 0.69±0.14 and 0.85±0.10 and most of the DSCs fell within the mean ± 2 standard deviations of the manual contours.

Conclusion: We developed a deep learning approach for automatic lung tumor segmentation and demonstrated the feasibility to achieve reasonably good segmentation.

Funding Support, Disclosures, and Conflict of Interest: This work was partially support by the CPRIT (Cancer Prevention Research Institute of Texas) grant RP110562-P2 and RP110562-C2.


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email