RT Journal Article
SR Electronic
T1 Consolidating Deep Learning Framework with Active Contour Model for Improved PET-CT Segmentation
JF Journal of Nuclear Medicine
JO J Nucl Med
FD Society of Nuclear Medicine
SP 1415
OP 1415
VO 62
IS supplement 1
A1 Yousefirizi, Fereshteh
A1 Rahmim, Arman
YR 2021
UL http://jnm.snmjournals.org/content/62/supplement_1/1415.abstract
AB 1415Objectives: Although convolutional neural networks (CNNs) have gained much attention in the last decade for segmentation, the limited capability of CNNs to capture small objects, fine boundaries and spatial details has been repeatedly reported. This issue needs to be addressed considering the importance of PET-CT segmentation of malignancy towards automated metabolic tumor volume measurements, radiomics analyses and radiotherapy planning. Our goal is to benefit from the classical intelligence of the active contour model (ACM) as an unsupervised approach for PET-CT segmentation in two ways: 1) Using the efficient energy minimization of ACM, i.e. incorporating Mumford-Shah (MS) functional as an additional term in the loss function of deep framework, to complement the semantic information used by CNNs. 2) Using the region-based active contour segmentation approach to refine object boundaries as a post-processing step. Methods: We analyzed 200 PET-CT images of head and neck cancer patients from four different institutions originating from study by Vallieres et al. (2017). We designed a deep framework for bi-modal PET-CT image segmentation based on generative adversarial network (GAN) inspired by SeGAN by Xue et al. (2018) with a multiscale nature designed for medical image segmentation that outperforms V-net. A modified V-net structure was used as the generator (G) network, and the architecture of the discriminator network was similar to the encoder part of the G network. First, the MS loss functional as proposed by Kim et al. (2019) is based on the spatial correlation in ground truth, compared to the cross-entropy as the loss function for most deep frameworks in PET-CT segmentation. Secondly, the segmented region by deep framework is used as the initial mask for the localized ACM to compute the energy function based on the region-based hybrid information of both PET and CT channels. Since the accuracy of ACM depends strongly on the number of iterations, the optimum number of iterations in our work was chosen based on the best DSC values (number of iterations=150). This helps preserve the automatic nature of the proposed segmentation technique. The segmentation results of (1) SeGAN, (2) SeGAN with MS loss (SeGAN(MS)), (3) SeGAN with post-processing ACM (SeGAN+ACM), (4) SeGAN with MS loss + post-processing ACM (SeGAN(MS)+ACM), and (5) V-net were evaluated by Dice (DSC), Jaccard coefficients (JSC) and Hausdorff distance (HD) criteria (±std). Paired t-test comparisons were performed for DSCs, JSCs and HD values between these 5 methods. Results: Mean DSCs, JSCs and HD values are presented in Table 1. Our results show that using both MS loss and post-processing step of ACM leads to superior performance (p-value <0.05). Using MS loss function depicts superior results compared to post-processing step of ACM (p-value<0.05). SeGAN showed lower performance (p-value <0.05) compared to integrating ACM in any of the proposed ways (p-value<0.05). All SeGAN implementations significantly outperformed V-net based segmentation. Table 1. Mean DSC, JSC and HD values (±std). (Stars (*) on values in a row denote significant difference compared to the following row). View this table: Conclusions: Segmentations within supervised CNN based framework are significantly improved by integrating ACM i.e. adding MS loss function and using ACM as the post-processing step. This idea can be applied to improve other CNN based frameworks i.e. V-net for PET-CT segmentation. It is worth mentioning that ACM post-processing step increases computation time for each patient from 20 seconds (inference time) in SeGAN(MS) up to three minutes in SeGAN(MS)+ACM, though with significant improvements.