Search

Ziel/Aim: Patients with locally advanced non-small-cell lung cancer (NSCLC) have a high risk of developing distant metastases. It has been shown that immunotherapy after radiochemotherapy can significantly improve the prognosis. Therefore, biomarkers for individualized therapy escalation are urgently needed. One such biomarker could be the total metabolic volume of primary tumor and lymph node (LN) metastases. However, delineation of LN metastases with currently available methods is time consuming and error-prone. The goal of this study was to investigate to which extend this delineation can be performed with deep learning methods. Methodik/Methods: Automated delineation was performed with a pretrained 3D U-Net convolutional neural network (CNN) previously derived for a different head and neck cancer delineation task. 517 [18F]FDG PET/CT scans of NSCLC patients were used for further network training and testing using a 5-fold cross-validation scheme. In these data, manual delineation and labeling of primary tumor and metastases was performed by an experienced physician serving as the ground truth for network training and testing. Ergebnisse/Results: The derived CNN models are capable of accurate delineation, achieving a Dice similarity coefficient of 0.854. Sensitivity of lesion detection was 0.841 and positive predictive value was 0.847. Accuracy of lesion classification as primary tumor or LN metastases was 82.2%. Schlussfolgerungen/Conclusions: In this work, we present a CNN able to perform delineation of and discrimination between primary tumor and lymph node metastases in NSCLC with only minimal manual corrections possibly required. It thus is able to accelerate study data evaluation in quantitative PET and does also have potential for clinical application.

Aim/Introduction: Patients with locally advanced non-small-cell lung cancer (NSCLC) have a high risk of developing distant metastases. It has been shown that applying immunotherapy after radiochemotherapy can significantly improve the prognosis for affected patients. In this context, biomarkers for individualized therapy escalation are urgently needed. One such biomarker could be the total metabolic volume of primary tumor and lymph node (LN) metastases (total tumor burden, TTB). However, delineation of tumor lesions with conventional methods is time consuming and error-prone, especially for the LN metastases. The goal of this study was to investigate feasibility of such delineation with deep learning methods. Materials and Methods: Automated delineation was performed with a 3D U-Net convolutional neural network (CNN) developed with the nnU-Net software package [1]. The default nnU-Net training parameters were modified to provide better training stability with small lesion targets as well as to better balance sensitivity vs. positive predictive value (PPV) of lesion detection. A dataset consisting of 517 [18F]FDG-PET/CT scans of NSCLC patients was used for the network training and testing following 5-fold cross-validation scheme. In these data, the ground truth labels were defined via manual delineation and labeling of primary tumor and metastases by an experienced physician. Results: The derived CNN models were capable of accurate delineation, achieving a mean (median) Dice similarity coefficient of 0.831 (0.891). The sensitivity and PPV of lesion detection was 0.974/0.829/0.887 and 0.963/0.741/0.824 for primary tumor/LN metastases/union of both, respectively. Accuracy of lesion classification as primary tumor or LN metastases was 92.1%. Manually and automatically derived TTBs were highly correlated with R2=0.96 and a mean absolute difference of 5.4 ml (after rejecting 1% of the outliers). Conclusion: In this work, we present CNN models able to perform delineation

MOTIVATION: Patients with locally advanced non-small-cell lung cancer (NSCLC) have a high risk of developing distant metastases. It was shown that immunotherapy after radiochemotherapy significantly improves the prognosis. Therefore, biomarkers to identify such patients are urgently needed. Here, we investigated the prognostic utility of total tumor burden (TTB) in NSCLC for prediction of distant metastases. METHODS: Altogether, 165 patients (65+/-9)y, 100m) with newly diagnosed NSCLC were included. All patients received FDG-PET/CT prior to definitive radiochemotherapy. In the PET images, the metabolically active volume (MTV) of the primary tumor and of all FDG avid lymph nodes was delineated with an adaptive threshold method. TTB was computed as the cumulative volume of primary tumor and lymph nodes. Survival analysis with respect to freedom from distant metastases (FFDM) was performed. RESULTS: Survival analysis revealed MTV and TTB as prognostic factors for FFDM (P=0.004 and P<0.001, respectively). Hazard ratio (HR) for TTB was significantly higher than HR for MTV (1.9 vs. 2.5, P=0.007). CONCLUSIONS: In the investigated group of patients, the inclusion of lymph nodes into MTV computation significantly increased the prognostic value of FDG-PET. Further investigations are necessary to confirm these preliminary results.