A convolutional neural network for fully automated total metabolic tumor volume delineation in patients with aggressive Non-Hodgkin lymphoma.

Abstract

PURPOSE: The [18F]FDG-PET-derived total metabolic tumor volume (TMTV) has a high prognostic value in patients with Hodgkin and Non-Hodgkin lymphoma. However, in order to enable TMTV as a biomarker for clinical use, an accurate and fast method of tumor delineation in lymphoma patients is needed. Deep-learning-based methods have shown promising results in this field and offer distinct advantages over classical approaches. Therefore, the goal of this work was to train a convolutional neural network (CNN) for delineation of all lymphoma lesions regardless of their size and uptake characteristics while performing the optimal contouring of each individual lesion. METHODS: A neural network was trained with the nnU-Net software package. A total of 1192 [18F]FDG-PET/CT scans from 716 patients with Non-Hodgkin lymphoma participating in the PETAL trial comprised the main dataset which was used for training. The ground truth delineation included all lesions that were clinically considered as lymphoma manifestations by an experienced observer and was developed iteratively with the assistance of intermediate CNN models. Performance of the trained network was assessed in the main dataset via 5-fold cross-validation as well as in external benchmark dataset (N = 60 scans). RESULTS: Comparing the manual and automated delineations in the main (external) dataset, the aggregated Dice coefficient reached 0.895 (0.715) and the corresponding TMTVs were highly correlated with R2 = 0.974 (0.767). The main (external) dataset contained a total of 8971 (713) manually delineated lesions, the detection sensitivity of which was 71.2% (77.6%) with the positive predictive value of 84.1% (63.1%). Univariate Cox regression analysis in the main dataset revealed both manually and automatically derived TMTVs as highly prognostic factors for progression-free survival with very similar hazard ratios (HR = 3.5; p < 0.001 and HR = 3.7; p < 0.001, respectively). CONCLUSION: In this study we presented a CNN model capable of accurate TMTV delineation in [18F]FDG-PET/CT images of lymphoma patients. It is trained to delineate all tumor lesions while accounting for typical caveats inherent in this task. The developed neural network allows for substantial acceleration of quantitative analysis of lymphoma imaging data and has the potential for supervised clinical use.