Factors Influencing the Accuracy and Coverage of CT-Based Lymph Node Delineation in Uterine Cervical Carcinoma: A Deep Learning Approach

Purpose To analyze the factors influencing the accuracy and errors in the delineation of draining lymph nodes for uterine cervical carcinoma using a CT imaging diagnostic system based on two convolutional neural networks—GoogLeNet and Faster R-CNN. Material and methods A total of 679 lymph nodes in 56 patients with Uterine Cervical Carcinoma (UCC) below the renal hilar level, around the main lumen vessels, and in the pelvic cavity were delineated by the image diagnostic system. Then, two associate chief physicians in the imaging department evaluated the outlined lymph nodes to check for any missed lymph nodes. The lymph nodes were categorized into the following groups based on their basic characteristics: 1. Size. Based on previous research regarding the capability of convolutional neural networks (CNN) to recognize nodular nodes and the likelihood of lymph nodes appearing in images, they were classified into three groups: those with a length and diameter of less than 0.5 cm, between 0.5 cm and 2 cm, and greater than 2 cm; 2. Location. Considering the probability of metastatic lymph nodes in cervical cancer, they were divided into bilateral pelvic wall, common iliac vessels, retroperitoneum (common location group), and pelvic lymph nodes (uncommon location group); 3. Presence or absence of necrosis. We analyzed the coverage rate of lymph nodes identified by the imaging diagnostic system and its consistency with manual diagnoses. We analyzed the coverage rate of lymph nodes identified by the imaging diagnostic system and its consistency with manual diagnoses. Results A total of 679 lymph nodes were identified in this study, of which 281 were smaller than 0.5 cm, and 265 were outlined by the imaging diagnostic system. There were 312 lymph nodes measuring between 0.5 cm and 2 cm, of which 298 were outlined by the imaging diagnostic system. A total of 86 lymph nodes were larger than 2 cm, with 80 outlined. The diagnostic coverage rates for the three groups were 94.31%, 95.51%, and 93.02%, respectively, showing no statistically significant difference. Among the lymph nodes, 587 were located in common areas, with 575 outlined, resulting in a coverage rate of 97.96%. There were 92 unusual lymph nodes, of which 68 were outlined, yielding a coverage rate of 73.91% (p ≤ 0.01), indicating a statistically significant difference. There were 32 necrotic lymph nodes, with 23 outlined, resulting in a coverage rate of 71.88%. Additionally, there were 647 lymph nodes without necrosis, and 620 were outlined, yielding a coverage rate of 95.83% (p ≤ 0.01), indicating a statistically significant difference. Conclusion The imaging diagnostic system for (UCC) based on convolutional neural networks demonstrates a high degree of consistency with manual diagnoses. The coverage rate of lymph node delineation is significantly influenced by the location of lymph nodes and the presence or absence of necrosis, whereas the size of lymph nodes does not have a significant effect on the diagnostic system’s coverage rate. Therefore, the recognition capability of the diagnostic system for lymph nodes with necrosis and those in unusual locations should be enhanced to improve the overall coverage rate of diagnosis.