Non-Invasive Prediction of 177 Lu-Dotatate Treatment Response Using 68 Ga-Dotatate PET/CT and Deep Learning
Abstract
Purpose
To develop and evaluate a custom convolutional neural network (CNN) using baseline 68Ga-DOTATATE PET/CT imaging to predict disease control following 177Lu-DOTATATE therapy in patients with neuroendocrine tumors (NETs).
Methods
We conducted a retrospective study using a subset of patients from a of a phase II, open-label, single-arm registry trial, consisting of 161 patients who received 177Lu-DOTATATE therapy until December 2022. Pre-treatment 68Ga-DOTATATE PET/CT scans and clinical data were collected to train a CNN based on ConvNeXt for binary classification of disease control. Prolonged disease control was defined as the absence of disease progression at 1.5 years following therapy. The dataset was split into 70% training (n=111), 15% validation (n=15), and 15% testing (n=15) stratified by disease control and sex. Organ volumes were segmented on CT scans using TotalSegmentator, an open-source AI tool. These anatomical segmentations were used in a custom-built segmentation algorithm, where liver-based thresholding was used to identify NETs. Tumor segmentations, PET/CT imaging, and clinical data, were used as inputs into the CNN. The model was trained for up to 500 epochs with binary cross entropy (BCE). Early stopping, applied with BCE and area under the receiver operating characteristic curve (AUC), was evaluated on the validation dataset. Accuracy, AUC, sensitivity, specificity, positive predictive value (PPV) and negative predictive value were evaluated on the testing dataset.
Results
Training ended at 490 epochs. The CNN demonstrated an accuracy of 75% (AUC 61%, sensitivity 71%, specificity 42%). However, the PPV is higher at 83%.
Conclusion
Predicting patient response to 177Lu-DOTATATE therapy using machine learning remains challenging. While demonstrating modest overall performance, the relatively high positive predictive value suggests that this approach may be useful for identifying patients likely to experience prolonged disease control. Such models could support clinical trial selection and provide added clinical reassurance when predictions are favorable.