Adaptive CNN Models for X-ray Scatter Correction in CBCT: A Transfer Learning Strategy

In recent years, convolutional neural networks (CNN) have demonstrated potential to perform xray scatter corrections, but there are still challenges to effectively use CNN: first, deep learning models benefit from large volumes of data. Second, it is not clear how generalizable these models are. In this work, a framework to address the previously described challenges by means of transfer learning (TL) and data simulation is proposed. The use of TL allowed adapting a pretrained CNN to specialized tasks, with relatively small volumes of data and in a fast manner. A data simulation approach enabled to easily generate tailored training datasets for given acquisition settings and anatomical regions. The pretraining was performed using geometric phantom data, while TL was achieved using anthropomorphic data. For the pretraining stage, it was observed that scatter ratio (SR) predictions were accurate (SSIM~0.8). This was further improved after TL, with a relatively short retraining time (~13 times faster than pretraining) and using considerably fewer samples compared to the pretraining dataset (~5 times smaller). In conclusion, a fast and low-cost methodology to generate task-specific CNN for scatter correction in CBCT was developed. CNN models trained with the proposed methodology were successful to correct x-ray scatter in anthropomorphic structures, unknown to the network, for simulated data.