Deep Neural Network-Assisted Computed Tomography Diagnosis of Metastatic Lymph Nodes from Gastric Cancer
Gastric cancer is a significant global health concern, ranking fifth among malignancies and third in tumor-related deaths. In China, it is particularly prevalent, with high mortality and poor prognosis. By the time of diagnosis, cancer cells have often infiltrated and metastasized, with lymph node metastasis being the primary form of spread. Accurate pre-operative assessment of lymph node metastasis is crucial for effective treatment, including surgery and neoadjuvant chemotherapy. However, traditional methods of evaluating lymph node metastasis through enhanced computed tomography (CT) scans are fraught with challenges, including technical difficulties and inevitable false-negative and false-positive results.
This study explores the application of deep neural networks, specifically Faster Region-Based Convolutional Neural Networks (FR-CNN), to improve the accuracy of CT diagnosis of perigastric metastatic lymph nodes (PGMLNs) in gastric cancer patients. The research aims to simulate the recognition of lymph nodes by radiologists and achieve more accurate identification results through deep learning.
The study involved a retrospective analysis of 750 gastric cancer patients who underwent CT scans at the Affiliated Hospital of Qingdao University from January 2011 to May 2018. Of these, 250 patients had post-operative pathological confirmation of PGMLNs. The research was approved by the hospital’s ethics committee and registered in the Chinese Clinical Trial Registry.
The methodology included two deep learning sessions: an initial learning phase and a precision learning phase. In the initial phase, 18,780 enhanced CT images and 1,371 labeled CT images from 313 patients were used for training. The precision learning phase involved 11,340 enhanced CT images and 1,004 labeled CT images from 189 patients. The labeled images were carefully examined and marked by experienced radiologists based on size, shape, and enhancement density values.
The FR-CNN was trained using these datasets to identify PGMLNs. The training process included two phases: initial training with 1371 labeled images and 18,780 original images, and precision training with 1004 labeled images and 11,340 original images. The network was initialized with pre-trained weights from ImageNet and fine-tuned using stochastic gradient descent and backpropagation.
The results demonstrated significant improvements in the accuracy of PGMLN identification. In the initial group, the mean average precision (mAP) value was 0.5019, and the area under the receiver operating characteristic curve (AUC) was 0.8995. After the precision learning phase, these metrics improved to 0.7801 and 0.9541, respectively. The FR-CNN achieved a recognition accuracy of 95.4%, significantly higher than traditional methods.
The study highlights several advantages of using FR-CNN for PGMLN recognition. The network can reconstruct lymph node distribution from multiple angles, improving detection rates and specificity. It also reduces false negatives caused by axial CT image limitations and false positives from structures like vascular cross-sections or perigastric adipose tissue. The FR-CNN provides comprehensive judgment on lymph node location, number, size, shape, and density, enhancing the accuracy of lymph node staging in gastric cancer.
The implications of this research are profound. Accurate pre-operative evaluation of PGMLNs is essential for determining the need for neoadjuvant chemotherapy, which has been shown to improve survival rates in advanced gastric cancer patients. The FR-CNN’s ability to quickly and accurately determine the number, locations, and diameters of metastatic lymph nodes can guide optimal treatment plans.
Additionally, the FR-CNN can assist in endoscopic submucosal dissection and lymph node dissection during radical gastrectomy. Accurate pre-operative understanding of PGMLN distribution and number is crucial for effective lymph node dissection, which is a critical factor in patient prognosis. The FR-CNN eliminates human error and fatigue, ensuring more accurate and efficient recognition of lymph nodes.
In conclusion, the FR-CNN represents a significant advancement in the pre-operative evaluation of PGMLNs in gastric cancer patients. By leveraging deep learning, the network achieves high judgment effectiveness and recognition accuracy, transforming the traditional manual strategy into an artificial intelligence-based approach. This innovation has the potential to improve diagnosis, treatment planning, and ultimately, patient outcomes in gastric cancer care.
doi.org/10.1097/CM9.0000000000000532
Was this helpful?
0 / 0