A convolutional neural network for automated delineation and classification of metabolic tumor volume in head and neck cancer in FDG-PET/CT

Aim: Image derived PET parameters such as metabolic tumor volume (MTV), total lesion glycolysis, and tumor asphericity of the primary tumor have been shown to be prognostic of clinical outcome of patients with head and neck cancer (HNC). Evaluation of lymph node metastases in addition to the primary tumor further increases the prognostic value of PET. Such analysis requires, however, accurate delineation and classification of all lesions which is very time-consuming when performed manually. The goal of this study is development of an automated tool for MTV delineation of primary tumor and lymph node metastases in HNC in PET/CT.

Methods: Automated delineation of the HNC cancer lesions was performed with a residual 3D U-Net convolutional neural network (CNN). 698 FDG PET/CT scans from 3 different sites and 4 public databases were used for network training (N=558) and testing (N=140). In these data, primary tumor and metastases were manually delineated and accordingly labeled by an experienced physician. This manual delineation served as the ground truth for network training. Performance of the trained network model was assessed in the test data using the Dice similarity coefficient for primary tumor, metastases, and the union of all lesions, respectively.

Results: The derived U-Net model is capable of accurate delineation of malignant lesions achieving a Dice coefficient of 0.847 for indiscriminative segmentation. Treating primary tumor and lymph node metastases as distinct classes yields Dice coefficients of 0.840 and 0.714 for the respective delineations.

Conclusions: In this work, we present the first CNN model for MTV delineation and classification in HNC. The developed network model allows to quickly perform satisfactory delineation of (and discrimination between) primary tumor and lymph node metastases in HNC with only minimal manual corrections possibly required. It thus is able to improve and to accelerate study data evaluation in quantitative PET and does also have potential for clinical application.