Click here to

Session: Imaging: CT Radiomics and Clinical Applications [Return to Session]

Multi-Scale Regional Fusion Based Interactive 3D Tumor Segmentation in Lung Cancer

R Wang1*, J Yang2, Z Mu3, Z Wang4, R Xu5, Z Zhou6, (1) Xidian University, ,,CN, (2) Xidian University, Xi'an, ,CN, (3) ,,,(4) Peking University Cancer Hospital & Institute, ,,CN, (5) Putian Unviersity, ,,(6) University of Central Missouri, Warrensburg, MISSOURI


SU-IePD-TRACK 2-7 (Sunday, 7/25/2021) 12:30 PM - 1:00 PM [Eastern Time (GMT-4)]

Purpose: Accurate lung tumor segmentation in computed tomography (CT) images is of great importance in diagnosis and target identification in lung cancer treatment. Most current segmentation methods require a large scale labeled data to train, while it is always difficult and time consuming to acquire labeled data. We aim to develop a new multi-scale regional fusion based interactive method for three-dimension (3D) tumor segmentation (MFIS).

Methods: Since no data is needed to train the model, seven patients collected from Peking University Cancer Hospital (Beijing, China) are used for testing. The manually contoured tumor by an experienced radiologist is used as ground truth. MFIS consists of three steps: (1) normalization. To reduce the potential influence of image acquisition protocols on tumor segmentation, CT images are normalized through min-max normalization. (2) 3D tumor block extraction through interactive information. The interactive markers such as a simple line are drawing on 2D images to indicate the tumor region. (3) Multi-scale regional fusion. Three phases are included in this step, they are multi-scale super-voxel pre-segmentation to divide the 3D image into a number of super-voxel regions in first phase. Then Super-voxel region merging is performed to merge the region with maximal similarity as the tumor. Finally, multi-scale average fusion is performed to improve segmentation accuracy through capturing both local and global contextual information. Dice coefficient (Dice), true positive rate (TPR) and false positive rate (FPR) are used for evaluation. The ideal segmentation is higher Dice and TPR, while lower FPR.

Results: The quantitative evaluation results for MFIS on Dice, TPR and FPR are 0.8821, 0.0052, and 0.8943, respectively.

Conclusion: A new multi-scale regional fusion based interactive method for three-dimension (3D) tumor segmentation (MFIS) in long cancer was developed. The experimental results demonstrated that MFIS can obtain better segmentation performance in a more efficient way.



    Not Applicable / None Entered.


    Not Applicable / None Entered.

    Contact Email