Predicting Lymph Node Malignancy with Convolutional Neural Networks: Insights from a Transatlantic and Multicenter Study in Endoscopic Ultrasound

Aims Endoscopic ultrasound (EUS) plays a crucial role in the evaluation of lymph nodes (LN), particularly in oncological settings where accurate LN assessment guides treatment by determining metastatic involvement. Malignant LNs typically exhibit features such as hypoechoic echotexture, sharply demarcated borders, rounded morphology, and size>10 mm, achieving a diagnostic accuracy of 80–100% when combined. However, these features co-occur in less than 30% of cases. EUS-guided fine-needle aspiration/biopsy (FNA/FNB) remains the gold standard for tissue diagnosis, with reported adequacy rates of 87–95%, but its success depends on the number and accessibility of target nodes. Artificial intelligence (AI), particularly convolutional neural networks (CNNs), offers promise in enhancing EUS-based LN characterization. AI could reduce unnecessary biopsies by enabling selective targeting and prioritize the most suspicious nodes for sampling in patients with multiple adenopathies. This multicenter study evaluates the effectiveness of a CNN in predicting LN malignancy using EUS images.

Methods This multicenter study included EUS images from four centers, three in Spain (Hospital Universitario Puerta de Hierro Majadahonda, Hospital Universitario Ramón y Cajal and Hospital Universitario Marqués de Valdecilla) and one in Brazil (Hospital das Clínicas da Faculdade de Medicina da Universidade de São Paulo). Lesions were labeled (“malignant” or “benign”) and delimited with bounding boxes. Definitive diagnoses were based on positive FNA/FNB or surgical specimens and, if negative, a minimum six-month clinical follow-up was required. A CNN was developed using the YOLO (“You Only Look Once”) model, including both detection and characterization modules. Performance metrics included detection rate (DR), sensitivity, specificity, accuracy, and AUC were calculated.

Results A total of 56270 images from 63 patients were used to develop the model. Patients were divided for constitution of training (n=39), validation (n=17) and test (n=7). The model had a DR for LNs of 92.2%. The CNN distinguished malignant from benign LNs with a sensitivity of 95%, a specificity of 88.8% and a precision of 90.8%. The overall accuracy for the distinction of both categories was 92.2%.

Conclusions To our knowledge, this study is one of the first to evaluate the performance of deep learning systems for LN assessment using EUS imaging, highlighting CNNs' potential to improve diagnostic accuracy. Preoperative LN status assessment is crucial for adjusting treatment plans based on metastatic involvement. Moreover, the model’s ability to predict benignity with high specificity could reduce the need for unnecessary follow-ups and biopsies. Our AI-powered imaging model shows excellent detecting and classification capabilities, highlighting its potential to provide a valuable tool to refine LN assessment with EUS, ultimately supporting more tailored and efficient management strategies for patients [1] [2] [3].

 

Scroll to top

Conflicts of Interest

One of the authors has participated as a speaker in Medtronic-sponsored eventsOne of the authors has participated in sponsored events by Boston Scientific.

• References

• 1 Chin YK, Iglesias-Garcia J, de la Iglesia D, Lariño-Noia J, Abdulkader-Nallib I, Lázare H, Rebolledo Olmedo S, Dominguez-Muñoz JE.. Accuracy of endoscopic ultrasound-guided tissue acquisition in the evaluation of lymph nodes enlargement in the absence of on-site pathologist. World J Gastroenterol 2017; 23 (31): 5755-63
• 2 Catalano MF, Sivak MV, Rice T, Gragg LA, Van Dam J.. Endosonographic features predictive of lymph node metastasis. Gastrointest Endosc 1994; 40 (04): 442-6
• 3 Tamanini G, Cominardi A, Brighi N, Fusaroli P, Lisotti A.. Endoscopic ultrasound assessment and tissue acquisition of mediastinal and abdominal lymph nodes. World J Gastrointest Oncol 2021; 13 (10): 1475-91

Publication History

Article published online:
27 March 2025