BLUE RIBBON POSTER RADIOPHARMACEUTICALS: Attention Med3D: A 3D Deep Learning Model Based on Transfer Learning and Attention Mechanism for Predicting Mid-Treatment Chemoradiation Response on FDG PET of La-NSCLC
Abstract
Purpose
Accurate assessment of early radiotherapy response in tumors provides crucial guidance for optimizing radiotherapy protocols. We developed a 3D deep learning model termed Attention Med3D based on transfer learning and attention mechanisms for predicting mid-treatment fluorodeoxyglucose (FDG) positron emission tomography (PET) response.
Methods
Twenty-five patients with locally advanced non-small cell lung cancer (LA-NSCLC) underwent baseline FDG-PET/CT imaging (PETpre) and repeat imaging during week 3 (PETmid) of chemoradiotherapy under the FLARE-RT trial (NCT02773238). PETpre and planned radiation dose (Dose) were used as inputs to predict voxel-wise SUV values in the 3D PETmid tumor volume. Data augmentation expanded training samples through rotation, flipping, and random scaling. Attention Med3D adopted an encoder-decoder architecture. The encoder was constructed through transfer learning from the Med3D pre-trained model for feature extraction. A decoder based on skip connections and attention mechanisms was designed to reconstruct spatial dimensions through upsampling and predict voxel-wise SUV values. To quantify treatment response, the relative SUV change (∆SUVratio) was calculated as ∆SUVratio = (SUVpre - SUVmid)/SUVpre. Model performance was evaluated through leave-one-patient-out cross-validation, with root mean square error (RMSE) and mean absolute error (MAE) calculated for both direct SUV prediction and the derived ∆SUVratio.
Results
Attention Med3D achieved tumor voxel-level SUV prediction with RMSE of 1.521 and MAE of 1.240. For the calculated ∆SUVratio, the model achieved an RMSE of 0.265 and MAE of 0.215. The model numerically outperformed 3D U-Net and 3D ResNet series models for both SUV prediction (RMSE: 1.610-1.810; MAE: 1.353-1.515) and calculated ∆SUVratio (RMSE: 0.298-0.346; MAE: 0.249-0.283), though without statistical significance (Friedman test, p > 0.05).
Conclusion
The proposed model leveraged transfer learning to construct a 3D deep learning framework suitable for limited-sample radiotherapy response prediction. This framework has the potential to help clinicians to optimize individualized radiotherapy protocols for lung cancer patients.