Precise Lymph Node Diagnosis for Esophageal Cancer on CT Images via User-Prior Guided Attentive Model

Abstract

<h3>Purpose/Objective(s)</h3> Esophageal lymph node (ELN) diagnosis is vital to ensure the effectiveness of radiation treatment but it is especially hard to achieve high diagnosis accuracy on CT images even for experienced doctors. Deep learning method has shown promising results in many automatic diagnosis tasks, but it has limited success on ELN diagnosis partially due to the low-contrast for the area on CT images. To improve the efficiency and accuracy of the clinical esophageal node diagnosis, we propose a novel semi-automatic lymph node diagnosis model incorporating example segmentation from users to attentively guide the determination. <h3>Materials/Methods</h3> A multi-site dataset with 144 patients was collected and used to demonstrate the method. Each patient has the original CT images, the segmentation labels of each node from experienced physicians, and the diagnosis results from the pathological reports. A total of 3550 nodes (2801 negative and 749 positive) from 2D slices were utilized. A user-prior guided attentive model was trained with inputting the images and example segmentation, which is based on the classical classifier, ResNet-50, and equipped with two user-prior guided attention modules. Specifically, a novel soft-attention (SAT) module was constructed to concentrate the network on the doctors' delineation to avoid misunderstanding. An extra node-relation exploring (NRE) module was designed to takes the image's features and delineation map of positive nodes as inputs and outputs the attentively enhanced features. The model performance was quantified using area under curve (AUC) after five-fold cross-validation. Most importantly, negative predictive value (NPV) was calculated to measure the accuracy of detecting negative nodes. To demonstrate the necessity of human interaction, a segmentation models (DeepLab-v3-plus) was trained to segment positive and negative node areas. An ablation study was performed to show the impacts of the SAT and NRE modules. To evaluate the model performance with the accuracy of example delineations, the labels were randomly expanded or corroded within ten pixels. <h3>Results</h3> The final model achieved 93.88% of AUC and 98.55% of NPV. The Dice Coefficients of the auto-segmented nodes were 16.32% and 53.45% respectively, indicating the challenges to automatically detect the positive nodes. For the ablation study, the AUC dropped to 86.78% removing NRE module and 56.88% after removing the two modules. The model performance was consistent with variable manual delineations, indicating that only coarse delineation is needed, and the model is robust in terms of users' inputs. <h3>Conclusion</h3> The proposed semi-automatic model can robustly and accurately assist the esophageal lymph node diagnosis on CT images. Our results demonstrated that only limited human inputs are needed and it can significantly improve the diagnosis accuracy. With further development, the method can be implemented into clinical workflow, providing precise diagnosis assistance in clinical practice.