Artificial Intelligence System of Faster Region-Based Convolutional Neural Network Surpassing Senior Radiologists in Evaluation of Metastatic Lymph Nodes of Rectal Cancer
Rectal cancer remains one of the most common gastrointestinal malignancies, with lymph node (LN) metastasis being a critical factor in determining patient prognosis and treatment strategies. Accurate evaluation of LN metastasis is essential for clinical decision-making, as it directly influences whether patients receive preoperative chemoradiotherapy or proceed directly to surgery. Magnetic resonance imaging (MRI) is widely used for LN assessment due to its high soft-tissue contrast. However, radiologists face challenges in accurately diagnosing LN metastasis due to the complexity of integrating factors such as LN shape, boundary, and signal intensity. Moreover, inter-observer variability and limited sensitivity in LN staging further complicate the diagnostic process. To address these issues, artificial intelligence (AI) systems, particularly deep learning models, have emerged as promising tools for improving diagnostic accuracy and efficiency. This study focuses on the development and validation of a Faster Region-Based Convolutional Neural Network (Faster R-CNN) AI system for the evaluation of metastatic LNs in rectal cancer patients.
Background and Significance
Rectal cancer is a leading cause of cancer-related mortality worldwide, with LN metastasis significantly increasing the risk of locoregional recurrence and poor prognosis. Accurate preoperative staging is crucial for determining the appropriate treatment regimen, which may include neoadjuvant chemoradiotherapy followed by surgery or immediate surgical intervention. MRI is the preferred imaging modality for preoperative staging due to its ability to provide detailed soft-tissue contrast. However, radiologists often struggle to accurately diagnose LN metastasis, leading to potential under- or over-staging. The development of AI systems, particularly those based on deep learning, offers a potential solution to these challenges by automating the detection and classification of metastatic LNs.
Study Design and Methods
This study involved 414 patients with rectal cancer who underwent radical resection between January 2013 and March 2015 across six clinical centers in China. The primary objective was to validate the accuracy of the Faster R-CNN AI system in diagnosing metastatic LNs, comparing its performance with that of senior radiologists and pathologists. The study included both methodological and clinical verification components. Methodological verification involved correlation analyses and consistency checks between the diagnoses made by Faster R-CNN, radiologists, and pathologists. Clinical verification focused on the long-term outcomes of patients, with a 36-month follow-up period to assess recurrence-free survival (RFS).
The Faster R-CNN system was trained using a dataset of MRI images annotated by radiologists. The system’s performance was evaluated using mean average precision (mAP), a metric that measures the accuracy of object detection in images. The MRI data included T2-weighted imaging (T2WI), fat-suppressed T2WI, and diffusion-weighted imaging (DWI) sequences, with specific parameters outlined in the study. The Faster R-CNN model was iteratively trained, and its performance was assessed after every 1,000 iterations.
Results
The study revealed significant correlations between the numbers of metastatic LNs diagnosed by Faster R-CNN, radiologists, and pathologists. The correlation coefficient between radiologists and Faster R-CNN was 0.912, indicating a high level of agreement. However, the correlation between pathologists and radiologists was lower (0.134), while the correlation between pathologists and Faster R-CNN was 0.448. These results suggest that Faster R-CNN diagnoses were more consistent with pathologists’ findings than those of radiologists.
In terms of N staging, the kappa coefficient between Faster R-CNN and pathologists was 0.573, compared to 0.473 between radiologists and pathologists. This indicates that Faster R-CNN was more accurate in N staging than radiologists. However, the system tended to overestimate the magnitude of N staging, particularly in patients classified as stage N2. For example, 72 patients (19.9%) classified as stage N1 by pathologists were misclassified as stage N2 by Faster R-CNN, while 10 patients (2.8%) classified as stage N2 were misclassified as stage N1.
The clinical verification component of the study involved a 36-month follow-up of 362 patients to assess RFS. Univariate survival analyses showed that sex, N staging based on pathologists, clinical staging, degree of tumor differentiation, and operation methods significantly influenced RFS. However, N staging based on radiologists and Faster R-CNN did not reach statistical significance in predicting RFS. Multivariate survival analyses confirmed that N staging based on pathologists, degree of tumor differentiation, and operation methods were independent predictors of RFS. Patients classified as stage N2 by pathologists had a significantly lower RFS rate (65%) compared to those classified as stage N2 by Faster R-CNN and radiologists (85%).
Discussion
The findings of this study demonstrate that the Faster R-CNN AI system outperforms radiologists in the evaluation of metastatic LNs in rectal cancer patients. The system’s high correlation with pathologists’ diagnoses and its ability to accurately classify N staging suggest that it could serve as a valuable tool for preoperative assessment. However, the system’s tendency to overestimate N staging, particularly in stage N2 patients, highlights the need for further refinement and validation.
The study also underscores the limitations of radiologists in diagnosing LN metastasis, as evidenced by the lower correlation with pathologists’ findings and the higher rate of misclassification. Faster R-CNN’s ability to process MRI images in just 20 seconds per case, compared to the average 600 seconds required by radiologists, further demonstrates its potential to improve diagnostic efficiency.
Despite these advantages, the study acknowledges that Faster R-CNN is not yet on par with pathologists in terms of diagnostic accuracy. Pathological examination remains the gold standard for LN assessment, but it is inherently limited by its post-operative nature. Faster R-CNN offers a preoperative alternative that, while not perfect, provides a significant improvement over radiologists’ assessments.
Conclusion
This study represents a significant step forward in the application of AI systems for the evaluation of metastatic LNs in rectal cancer patients. The Faster R-CNN AI system demonstrates superior accuracy and efficiency compared to radiologists, offering a promising tool for preoperative staging. However, further research is needed to address the system’s limitations and to validate its performance in larger, more diverse patient populations. As AI technology continues to evolve, it has the potential to revolutionize the field of oncology by enhancing diagnostic accuracy, improving patient outcomes, and reducing the burden on healthcare professionals.
doi.org/10.1097/CM9.0000000000000095
Was this helpful?
0 / 0