Prediction Error Propagation: A Novel Strategy to Enhance Performance of Deep Learning Models in Seminal Segmentation
Reza Karimzadeh, Emad Fatemizadeh, Hossein Arabi, Habib Zaidi · 2021 IEEE Nuclear Science Symposium and Medical Imaging Conference (NSS/MIC) · 2021
Segmentation of tumor and healthy surrounding organs, known as organs at risk (OAR), plays a key role in accurate diagnosis and treatment planning. To this end, many automated methods have been proposed, especially using deep learning methods. Convolutional neural networks (CNNs), inspired by the human visual cortex, entail the feedforward connections, however, the feedback connections from higher levels/layers to lower levels/layers in the visual cortex are missing in these models. These feedback connections are intended to carry the prediction errors of incoming information for the next moment/step to minimize the overall errors between reality and the prediction in the visual cortex. Inspired by the visual cortex structures, these feedback connections were proposed to be considered into CNN models in the form of prediction error propagation (PEP). To this end, a novel framework has been proposed that consists of three parts; encoder, reconstruction, and segmentation parts, wherein the input image is passed through encoder and reconstruction parts to create an error map of the predicted information/images (similar to what happens in visual cortex). This prediction error map is then propagated/feedbacked to the encoder part to aid accurate segmentation. The proposed PEP framework was implemented on a conventional Unet model (PEP-Unet). The PEP-Unet was trained for a multi-organ segmentation task from CT images (SegTHOR public dataset). The performance of the PEP-Unet was compared with the conventional Unet architecture using the standard segmentation indices. The proposed PEP framework led to an overall 11% improvement in terms of the Dice index. This study demonstrated the effectiveness of the PEP framework for deep learning-based seminal segmentation.