Deep Learning for Automatic Prediction of Lymph Node Station Metastasis in Esophageal Cancer Patients from Contrast-Enhanced CT

Abstract

The diagnosis of lymph node (LN) metastasis in computed tomography (CT) is an essential yet challenging task in esophageal cancer staging and treatment planning. Although criteria (e.g., RECIST, morphological/texture features) are proposed to predict LN metastasis, the diagnostic accuracy remains low with sensitivity <50% and specificity <75%, as reported in previous studies. Deep learning (DL) has the potential to address this issue by learning from large-scale labeled data. However, due to the practical surgery procedure in lymph node dissection, it is difficult to pair the metastasis of individual LN reported in the pathology report to the LN instance found in the CT image. Hence, in this study, we first use pathology reports to determine the LNS metastasis, then develop a multiple instance deep learning (MIDL) model to predict lymph node station (LNS) metastasis. We collected 1200 esophageal cancer patients with preoperative contrast-enhanced CT before surgery. A recently developed automatic mediastinal LNS segmentation model was first applied to segment LNS of 1 to 8 based on the IASLC protocol. For each LNS, the local CT region of interest (ROI) was cropped to generate a station-wise CT patch, where the LNS was labeled as metastatic if at least one metastatic LN was indicated in the pathology report. Using the station-wise CT patch and LNS label, we train a 3D MIDL model, MobileNetV3, to predict LNS metastasis. To better provide the LN position priors in MIDL, LN instances (with a short axis >4mm) were also segmented using an automatic LN detection algorithm and were added to the MIDL model as an auxiliary input. Five-fold cross-validation was conducted to evaluate the MIDL performance. The MIDL model's performance is summarized in Table 1. The MIDL model incorporating an additional LN instance mask demonstrated a superior overall AUC of 0.7539, surpassing the model without the LN mask input by 2.93%. The specificity was evaluated at a threshold resulting in a recall of 0.7, and the best model outperformed the CT input model in terms of specificity by 2.11%. This highlights the value of including the LN position prior to the MIDL model. Notably, when a threshold was set to result in a specificity of 75%, the best MIDL model demonstrated a significantly higher recall compared to the previously reported clinical diagnostic recall (39.7% vs. 63.21%). We developed a MIDL classification model to predict LNS metastasis using CT scans of 1200 patients. Our findings suggest that the MIDL model can substantially improve LNS metastasis prediction and has the potential to play an essential role in cancer staging, treatment planning, and prognostic analysis.