Abstract
1427
Objectives: Lymphoma lesion segmentation is an important step to compute total metabolic tumor volume (TMTV) and other radiomics features towards improved predictive modeling. Automatic segmentation of diffuse large B-cell lymphoma (DLBCL) lesions on [18F]FDG PET images is difficult due to: (1) inherent partial volume effect, high noise and low-resolution characteristics of PET images & heterogeneities in lesion uptake; and (2) high uptake values of normal organs, radiopharmaceutical clearance in organs, and diversity of lymphoma lesion sites. There is no consensus on an optimal thresholding method for the segmentation of DLBCL lesions (Cottereau et al. 2018). Deep learning frameworks (DLF) require laborious manual delineations to create reliable training datasets, with inter-observer and inter-center variabilities. Here, we aim at developing a new semi-automatic segmentation method. For this, we used improved fuzzy c-means clustering (FCM) and localized active contour model (LACM) to consider the potential of conventional segmentation techniques (CST) to reach deep framework performances.
Methods: The study population included interim [18F]FDG-PET images of 30 DLBCL patients. Two experienced nuclear medicine physicians performed manual lesion segmentation, considered as the ground truth (N=91 lesions). To mitigate the high sensitivity of FCM to noise, penalty term consisting of Euclidean distance between pixels and their neighborhood was added to the objective function. To tackle the time-consuming task of integrating the spatial information, the fast and robust FCM (FRFCM) based on morphological reconstruction & membership filtering was introduced. Although FRFCM is capable of boundary preservation, it is prone to leakage for segmentation. Previous studies applied FCM and ACM within region of interests (ROIs) initially defined by manual delineations in transaxial slices to prevent leakage. To automatically suppress leakage, we include 3 steps: (1) FRFCM segmentation is applied on maximum intensity projection (MIP) images first; normal organ areas are manually excluded at this stage. (2) FRFCM is applied on transaxial slices of the corresponding segmented regions. (3) LACM further refines the segmentations. The parameters of FRFCM and LACM were kept the same for all data. Dice similarity (DSC), Jaccard coefficients (JAC) and Hausdorff distance (HD) measures were calculated to assess performance and reported by median & interquartile ranges (IQR = IQ25-IQ75) to compare with other studies.
Results: Segmentation performance on MIP images reached a mean DSC value of 0.84 ± 0.12 after normal organ removal. FRFCM showed a median DSC 0.57 [IQR: 0.55-0.58], median JAC 0.54 [IQR:0.46-0.58], and median HD 4.75 [IQR: 4.03-6.09]. ACM method had median DSC 0.66 [IQR:0.58 to 0.67], median JAC 0.63 [IQR: 0.57-0.66], and median HD 3.84 [IQR: 3.52-4.35]. Recent deep framework studies were reported for DLBCL: nnU-net on PET/CT by Blanc-Durand et al. 2020 (median DSC 0.79 [IQR: 0.66-0.87]) and Deepmedic on PET by Weisman et al. 2020 (median DSC=0.64 [IQR: 0.41-0.77]). Our results are not far from these studies, pending implementation of an automatic framework for the removal of normal organ and improvement of the segmentation performance
Conclusions: Our FRFCM/ACM segmentation method does not require pre-specified ROIs on the transaxial slices, and maintains unsupervised ease of use of conventional methods. The only manual intervention is the elimination of normal organs that has been also pursued manually in some proposed end-to-end DLF for DLBCL segmentation. In future work, normal organs can be excluded by atlas-based segmentation on CT images. Our study shows that there is room for improvements in CST to achieve performances nearly on par with DLF. Our semi-automatic proposed method also has the potential to overcome limitations of thresholding techniques with under/overestimation of TMTV.