Chinese Guideline for the Application of Rectal Cancer Staging Recognition Systems Based on Artificial Intelligence Platforms (2021 Edition)
Imaging evaluation has become a cornerstone in the treatment planning for rectal cancer, whether for surgical intervention or neoadjuvant chemoradiotherapy. However, the interpretation of imaging results relies heavily on experienced radiologists, and the shortage and uneven distribution of such personnel often lead to delays and biases in diagnosis. To address this issue, the development of automatic recognition systems for rectal cancer staging based on artificial intelligence (AI) platforms has been proposed. These systems aim to partially replace the manual work of practitioners, enabling rapid and accurate identification of rectal cancer staging.
The guideline focuses on the use of faster region-based convolutional neural networks (FR-CNNs) to learn and train from a large number of images. These networks are designed to identify and label lesions, automatically delineate target areas, and perform three-dimensional reconstructions. The pre-operative evaluation of rectal cancer is primarily based on the tumor-node-metastasis (TNM) staging system and the circumferential resection margin (CRM). The guideline emphasizes the use of AI recognition systems to evaluate four key parameters: T staging, N staging, CRM, and extramural vascular invasion (EMVI).
Establishment of the Image Database
The image database was established using magnetic resonance imaging (MRI) images acquired from 3.0T MRI scanners, including those from GE, Siemens, and Philips. The main scanning sequences and parameters are detailed in Supplementary Table 1. After image collection, sequences with diagnostic value were selected and input into a deep neural network (DNN) to build the database.
Identification of Parameters in the Training Set of the AI Recognition System
The expert committee defined the identification of four parameters in the training set. T staging was determined based on the TNM staging system of the American Joint Committee on Cancer and Horvat et al’s study. T1 tumors infiltrate the submucosa, T2 tumors extend into the muscularis propria, T3 tumors show a discontinuity of the muscularis propria with extension into the mesorectum, and T4 tumors infiltrate the peritoneal reflection or other pelvic organs and structures. For N staging, the size, shape, and signal of lymph nodes (LNs) were used as diagnostic criteria. A lymph node was considered positive if the smallest diameter on T2-weighted imaging (T2WI) of MRI was ≥5 mm, the shape was irregular with indistinct edges, dispersion-enhanced imaging showed a high signal, or there were upper and lower discontinuities between image layers. Positive CRM was defined as a distance of ≤1 mm between the outer edge of the tumor and the mesenteric fascia, based on the criteria of the MERCURY Research Group. For EMVI, the scoring system from Smith et al’s study was used. On MRI T2WI, moderate signal intensity in the blood vessels with slightly enlarged contour and caliber, or clear tumor signal with obviously irregular or nodular dilatation of the vascular shape, was considered positive.
Identification Process
Two senior imaging experts and one colorectal surgery expert jointly evaluated and marked the correlation factors of rectal cancer staging on MRI images. In cases of inconsistency, a fourth imaging expert was consulted to reach a final conclusion. Approximately three marked images with diagnostic value were selected for each patient and input into the DNN to establish the MRI image database.
FR-CNN Training Framework
The FR-CNN framework was used to achieve the automatic detection of rectal cancer staging parameters. The system consists of two modules: a region proposal network and a module based on the candidate areas of the FR-CNN target detector. These two modules were combined into a unified target detection network, sharing the same convolution layer. The bounding box of suspected lesions served as the image output, with parameters of the output probability scores.
Selection of Clinical Validation Evaluation Tools
Receiver operating characteristic (ROC) and precision-recall (PR) curves were used as evaluation tools for the clinical verification of the AI automatic recognition system. Both ROC and PR are excellent evaluation tools in the field of detection and classification involving DNNs. The area under the curve (AUC) value, calculated based on the ROC curve, was used as the final evaluation indicator.
Result Interpretation of AI Recognition System
The AI recognition system, through deep learning and verification, aims to achieve higher accuracy evaluation indexes. ROC curve evaluation is recommended, and when the AUC ≥90%, the AI recognition system is considered ready for clinical application. The system locates and calculates probabilities for the four indicators of T staging, N staging, CRM, and EMVI of rectal cancer, providing a reference for clinicians in the judgment of pre-operative staging and the formulation of diagnosis and treatment plans. The guideline divides the probability into three levels: highly reliable, possible compliance, and poor compliance. Highly reliable includes T staging ≥90%, N staging, CRM, or EMVI ≥80%; possible compliance includes 90% > T staging ≥ 70%, 80% > N staging, CRM, or EMVI ≥ 60%; and poor compliance includes T staging < 70%, N staging, CRM, or EMVI < 60%.
Clinical Application Scenario of AI Recognition System
The guideline recommends the use of high-resolution MRI images of patients with rectal cancer, with the results input into the AI recognition system to obtain the identification results of T staging, N staging, CRM, and EMVI. For T1N0 early colorectal cancer patients, transanal local resection, transanal endoscopic microsurgery, or transanal minimally invasive surgery can be chosen. For pre-operative evaluation results that do not meet the above conditions, including the possibility of lateral LN metastasis, pre-operative neoadjuvant chemoradiation is recommended, followed by secondary identification using the AI recognition system. To achieve clinical complete response (cCR), transanal local resection or a watch-and-wait strategy is then recommended. For patients who have not reached cCR, radical total mesorectal excision (TME) surgery is recommended. If secondary identification still identifies lateral LN metastasis, selective lateral LN dissection is recommended and performed.
Conclusion
The development and application of AI recognition systems in rectal cancer staging aim to address the challenges posed by the shortage and uneven distribution of experienced radiologists. By leveraging FR-CNNs and deep learning, these systems provide rapid and accurate identification of key staging parameters, supporting clinicians in the formulation of diagnosis and treatment plans. The guideline emphasizes the importance of high-resolution MRI images and the use of ROC curve evaluation to ensure the reliability of the AI recognition system in clinical practice.
doi.org/10.1097/CM9.0000000000001483
Was this helpful?
0 / 0