Habitat radiomics assists radiologists in accurately diagnosing lymph node metastasis of adenocarcinoma of the esophagogastric junction

Abstract

Objectives: This study aimed to develop a habitat radiomics (HR) model capable of preoperatively predicting lymph node metastasis (LNM) in adenocarcinoma of the esophagogastric junction (AEG) and to implement its use in clinical practice.

Methods: In this retrospective analysis, 337 patients from three centers were enrolled and divided into three cohorts: training, validation, and test (208, 52, and 77 patients, respectively). We constructed HR models, conventional radiomics models, and combined models to identify LNM in AEG. The area under the curve (AUC) was employed to identify the optimal model, which was then evaluated for assisting radiologists in the empirical and RADS groups in diagnosing LNM. Finally, the prediction process of the optimal model was visualized using SHAP plots.

Results: The HR model demonstrated superior performance, achieving the highest AUC values of 0.876, 0.869, and 0.795 in the training, validation, and test cohorts, respectively. Regardless of seniority, the empirical group of radiologists showed a significant improvement in the AUC and accuracy when using the HR model, compared to working alone (p < 0.05). Furthermore, the RADS group radiologists exhibited strong reclassification ability, effectively reevaluating patients with false-negative LN initially classified as Node-RADS score 1 or 2 by themselves.

Conclusion: The HR model facilitates the accurate prediction of LNM in AEG and holds potential as a valuable tool to augment radiologists' diagnostic capabilities in daily clinical practice.

Critical relevance statement: The habitat radiomics model could accurately predict the lymph node status of adenocarcinoma in the esophagogastric junction and assist radiologists in improving diagnostic efficacy, which lays the foundation for accurate staging and effective treatment.

Key points: Accurate lymph node diagnosis in esophagogastric junction adenocarcinoma is beneficial for prognosis. Habitat radiomics model accurately predicted and assisted physicians in diagnosing lymph nodes. The habitat model effectively reclassified false-negative lymph nodes at Node-RADS 1 and 2.

Keywords: Adenocarcinoma of the esophagogastric junction; Computed tomography; Habitat radiomics; Lymph node; Node-RADS.

Fig. 1

Flowchart of the research. M, model

Fig. 2

Model performance comparison. Legend: Receiver operating characteristic (ROC) curves and Delong test in the training cohort (a, d), validation cohort (b, e), and test cohort (c, f). HR, habitat radiomics; CR, conventional radiomics; AUC, the area under the curve

Fig. 3

Comparisons of diagnostic performance between habitat radiomics (HR) models and radiologists. Legend: The figure shows the variation in radiologists before and after using the HR model to aid diagnosis of lymphatic node metastasis (LNM) in the validation cohort (a, b) and the test cohort (c, d)

Fig. 4

Sankey diagram for Node-RADS. Legend: a, c, e, g present Sankey plots of Node-RADS scores for LN assessed by RADS group physicians Junior 3 and Senior 3 in the validation and test cohorts. Lighter pink lines indicate correct diagnoses, while darker pink lines denote incorrect diagnoses. b, d, f, h are based on the independent physician scores mentioned earlier, incorporating the HR model as an intermediate node. Reclassified cases are represented by different yellow lines: earthy yellow denotes cases reclassified as positive in Node-RADS scores 1–2, while orange represents those reclassified as negative in scores 3–5. +, LNM positive; -, LNM negative

Fig. 5

Line plot of diagnostic metrics at different scores for Node-RADS. a Accuracy. b Sensitivity. c Specificity. d PPV. e NPV

Fig. 6

SHAP visualization of HR models. Legend: a SHAP bar chart. b SHAP summary plot. c SHAP heatmap. d SHAP stacked force plot. SHAP, Shapley additive explanation