Lymphoma segmentation from 3D PET-CT images using a deep evidential network

Abstract

An automatic evidential segmentation method based on Dempster-Shafer theory and deep learning is proposed to segment lymphomas from three-dimensional Positron Emission Tomography (PET) and Computed Tomography (CT) images. The architecture is composed of a deep feature-extraction module and an evidential layer. The feature extraction module uses an encoder-decoder framework to extract semantic feature vectors from 3D inputs. The evidential layer then uses prototypes in the feature space to compute a belief function at each voxel quantifying the uncertainty about the presence or absence of a lymphoma at this location. Two evidential layers are compared, based on different ways of using distances to prototypes for computing mass functions. The whole model is trained end-to-end by minimizing the Dice loss function. The proposed combination of deep feature extraction and evidential segmentation is shown to outperform the baseline UNet model as well as three other state-of-the-art models on a dataset of 173 patients.

Application of belief functions to medical image segmentation: A review

“…One evidence comes from a probabilistic model, and the other comes from an evidential model. Paper (Huang et al, 2021b) and (Huang et al, 2022) confirmed the similarity of ENN and RBF when acting as an evidential classifier and applied both the two models within a deep neural network for lymphoma segmentation. We think more achievements could be obtained by cooperating BFT and popular deep medical image segmentation model, such as UNet, VNet (Milletari et al, 2016) and nnUNet (Isensee et al, 2018), in the future.…”

“…Radial basis function (RBF) network. As confirmed in (Huang et al, 2022) that RBF network can be an alternative approach of ENN that based on the aggregation of weights of evidence.…”

“…The idea of applying the RBF within convolutional neural network (CNN) was first proposed by (Huang et al, 2022). In this paper, the RBF module becomes a "evidential module" that is plugged into the output of a CNN instead of the usual Softmax layer.…”

“…The reported quantitative and qualitative results showed that the proposal outperforms the state-of-the-art methods. Based on the first work, Huang et al (2022) first compared the performance of RBF and ENN and then applied them within the deep neural network (UNet) for lymphoma segmentation. The segmentation performance confirmed that RBF is an alternative approach of ENN (Denoeux, 2019b) to act as an evidential classifier and showed that the proposal outperforms the baseline method UNet and the state-of-the-art.…”

“…Discounting (Shafer, 1976) is an effective operation for multiple evidence fusion when counter with imperfect or unreliable evidence. For example, Huang et al (2022) claimed that mass function m comes from each prototype could be seen as a discounted Bayesian mass function with a discount rate in ENN. Thus the most reliable mass functions can be obtained after fusing the evidence from different prototypes.…”