Deep cross-modality (MR-CT) educed distillation learning for cone beam CT lung tumor segmentation

Jue Jiang,Sadegh Riyahi Alam,Ishita Chen,Perry Zhang,Andreas Rimner,Joseph O. Deasy,Harini Veeraraghavan
DOI: https://doi.org/10.1002/mp.14902
2021-04-21
Abstract:Despite the widespread availability of in-treatment room cone beam computed tomography (CBCT) imaging, due to the lack of reliable segmentation methods, CBCT is only used for gross set up corrections in lung radiotherapies. Accurate and reliable auto-segmentation tools could potentiate volumetric response assessment and geometry-guided adaptive radiation therapies. Therefore, we developed a new deep learning CBCT lung tumor segmentation method. Methods: The key idea of our approach called cross modality educed distillation (CMEDL) is to use magnetic resonance imaging (MRI) to guide a CBCT segmentation network training to extract more informative features during training. We accomplish this by training an end-to-end network comprised of unpaired domain adaptation (UDA) and cross-domain segmentation distillation networks (SDN) using unpaired CBCT and MRI datasets. Feature distillation regularizes the student network to extract CBCT features that match the statistical distribution of MRI features extracted by the teacher network and obtain better differentiation of tumor from background.} We also compared against an alternative framework that used UDA with MR segmentation network, whereby segmentation was done on the synthesized pseudo MRI representation. All networks were trained with 216 weekly CBCTs and 82 T2-weighted turbo spin echo MRI acquired from different patient cohorts. Validation was done on 20 weekly CBCTs from patients not used in training. Independent testing was done on 38 weekly CBCTs from patients not used in training or validation. Segmentation accuracy was measured using surface Dice similarity coefficient (SDSC) and Hausdroff distance at 95th percentile (HD95) metrics.
Image and Video Processing,Computer Vision and Pattern Recognition
What problem does this paper attempt to address?
The problem that this paper attempts to solve is to achieve accurate automatic segmentation of lung tumors in cone - beam computed tomography (CBCT) images. Specifically, although CBCT imaging is widely used in the treatment room, due to the lack of reliable segmentation methods, CBCT is only used for rough position correction. If an accurate and reliable automatic segmentation tool can be developed, it will be helpful for evaluating volume response and implementing geometric - guided adaptive radiotherapy. Therefore, the researchers have developed a new CBCT lung tumor segmentation method based on deep learning. ### Research Background and Objectives - **Background**: Lung cancer is the leading cause of cancer - related deaths among men and women in the United States. For inoperable or unresectable locally advanced non - small - cell lung cancer (LA - NSCLC), the standard treatment method is radical radiotherapy. However, when planning and implementing high - dose treatment, accurately delineating the target tumor and normal organs is a key technical challenge. - **Problem**: Although CBCT imaging is available, its soft - tissue contrast is low, making it difficult to extract features that clearly distinguish the target and background structures, even when using deep - learning methods. This limits the clinical application of CBCT, especially in geometric correction required for adaptive radiotherapy. ### Method Overview The researchers proposed a cross - modal - induced distillation (CMEDL) method, using magnetic resonance imaging (MRI) to guide the training of the CBCT segmentation network, thereby extracting more informative features. Specific methods include: 1. **Unpaired domain adaptation (UDA)**: Use a generative adversarial network (GAN) to synthesize pseudo - MRI images, which are used to train the segmentation network. 2. **Cross - domain segmentation distillation network (SDN)**: It includes a teacher MRI segmentation network and a student CBCT segmentation network. Through feature distillation, the statistical distribution of the CBCT features extracted by the student network is made to match the MRI features extracted by the teacher network, thereby improving the ability to distinguish tumors from the background. ### Experimental Results - **Performance improvement**: The CMEDL method significantly improves the segmentation accuracy of Unet and DenseFCN networks. Compared with 2D and 3D networks using only CBCT, the CMEDL method shows better performance in terms of the surface Dice similarity coefficient (SDSC) and 95% Hausdorff distance (HD95) metrics. - **Comparative experiment**: The CMEDL method also outperforms alternative frameworks using UDA and MRI segmentation networks. ### Conclusion The research results show that the introduced CMEDL method can generate relatively accurate lung cancer segmentation from CBCT images. Future verification on larger data sets is required to achieve clinical translation. ### Keywords - CBCT segmentation - Lung tumor - MRI - guided segmentation - Distillation learning - Adversarial deep learning Through this method, the researchers hope to more widely apply CBCT images in clinical practice, especially in adaptive radiotherapy, and improve the accuracy and effectiveness of treatment.