Deep learning model for the prediction of lymph node metastasis in esophageal squamous cell carcinoma using MIP FDG-PET images: a retrospective validation study

PURPOSE: Accurate preoperative diagnosis of lymph node (LN) metastasis in esophageal squamous cell carcinoma (ESCC) is crucial for determining treatment strategies, including organ-sparing therapies. We aimed to develop and validate a deep learning (DL) model using rotational maximum-intensity projection (MIP) 18 F-FDG PET images to improve the prediction of LN metastasis. MATERIALS AND METHODS: This retrospective study included 185 patients with ESCC (146 receiving neoadjuvant chemotherapy) who underwent preoperative imaging using a silicon photomultiplier PET scanner. A convolutional neural network (CNN) was developed using six rotational MIP images (angles from - 60° to 90°). The architecture utilized a ResNet-50 backbone with weight sharing across views. The ground truth for LN metastasis was established via postoperative histopathology, including Grade 3 pathological response as positive. The performance of the model in the test set (n = 36) was compared with that of radiologist reports and SUVmax analysis using the area under the receiver operating characteristic curve (AUC) and diagnostic accuracy. RESULTS: The CNN model achieved an AUC of 0.82 (95% CI 0.57-0.94), which was higher but not significantly different from the SUVmax method (0.77, p > 0.05). Although the differences were not statistically significant (p > 0.05), the CNN model achieved the highest absolute diagnostic accuracy of 86% at the optimal threshold, compared with the SUVmax-based method (67%), the clinical data model (75%), and radiologist reports (69%). Notably, while the CNN model exhibited a lower specificity (75%) than radiologist reports (92%), it demonstrated high sensitivity (92%) compared to radiologist (58%) and SUVmax (63%), effectively serving as a diagnostic safety net. CONCLUSION: Our PET-based CNN model using rotational MIP images demonstrated comparable diagnostic accuracy and higher sensitivity for LN metastasis in ESCC than the conventional methods. This proof-of-concept highlights its potential as a preoperative tool for optimizing treatment selection and minimizing understaging.