CT-Based Radiomics Can Improve the Prediction of Recurrent Prostate Cancer-Positive Lymph Nodes

Abstract

In patients with prostate cancer (PC), PSMA-PET/CT is currently the most preferred method for lymph node (LN) staging in certain regions. Globally, CT-based LN evaluation using the short diameter axis is still often used to determine the node status. Optimization of treatment planning according to the affected LNs is crucial for a successful radiation therapy (RT). The purpose of this retrospective study is to develop CT-based radiomic models for an optimized CT-based LN classification using LNs from recurrent PC patients who underwent PSMA-radio-guided surgery with confirmed histology. A total of 160 LNs with confirmed histology (114 positive and 46 negative) from a retrospective cohort of 87 recurrent PC patients were randomized to a training cohort (105 LNs) and a testing cohort (55 LNs). Suspicious LNs on (68Ga-PSMA-11)-PET/CT and adjacent non-diseased templates, were selectively surgically resected using a gamma probe following intravenous application of radioactively labeled PSMA with 99mTechnetium (73 patients) and 111Indium (14 patients). Histologically confirmed LNs were manually segmented using contrast-enhanced diagnostic CT datasets. 86 radiomic parameters including shape, intensity, and local binary pattern (LBP) features were extracted using a software package in python. On the training cohort, the least absolute shrinkage and selection operator (LASSO) algorithm was used to develop three radiomic models: "shape/intensity,” "LBP,” and a combined model. Logistic regression models were developed for clinical parameters including LN short diameter, LN volume, and a CT-based expert evaluation (CT-expert rating). Prediction performance stability was evaluated by bootstrapping (1000 fold). The resulting models were validated using the testing cohort. The LBP radiomic model showed the best prediction performance (area under the curve (AUC)-Training: 0.870, AUC-Testing: 0.836) in comparison to the shape/intensity radiomic model (AUC-Training: 0.838, AUC-Testing: 0.794). The combined radiomic model did not increase the discriminative performance further. Clinical parameters showed overall a lower predictive performance. Clinically, LN short diameter had the highest AUC values (AUC-Training: 0.750, AUC-Testing: 0.732) followed by LN volume (AUC-Training: 0.720, AUC-Testing: 0.571) and CT-expert rating (AUC-Training: 0.650, AUC-Testing: 0.644). The LBP radiomics model performed significantly better than LN short diameter (p=0.0029), LN volume (p=0.0285) and CT-expert rating (p<0.00001). We first developed a CT-based radiomic model to improve the detection of metastatic LNs in recurrent PC patients. The model was successfully internally validated. Such a model could be used to improve radiation therapy planning in case of unavailability of PSMA-PET imaging techniques.