📄 medicine

Functional Agent-Driven Computational Anatomy for Preoperative Siewert Classification of Adenocarcinoma of the Esophagogastric Junction Using Three-Dimensional Contrast- Enhanced CT: a retrospective multicenter diagnostic accuracy study

This retrospective multicenter study demonstrates that FACET, a functional agent-driven computational anatomy framework, improves the reproducibility and accuracy of preoperative Siewert classification for esophagogastric junction adenocarcinoma by converting subjective imaging assessments into calibrated, traceable 3D distance measurements on contrast-enhanced CT.

Original authors: Yiting Liu, Zifan Chen, Ruojin Li, Shuangxi Li, Li Zhang, Xiaoting Li, Yinkui Wang, Kan Xue, Fei Shan, Lei Tang, Ziyu Li

Published 2026-10-14

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yiting Liu, Zifan Chen, Ruojin Li, Shuangxi Li, Li Zhang, Xiaoting Li, Yinkui Wang, Kan Xue, Fei Shan, Lei Tang, Ziyu Li

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the complex landscape of human anatomy, few boundaries are as critical to surgical planning as the junction where the food pipe meets the stomach. This is the esophagogastric junction, a narrow passageway where a specific type of cancer, known as adenocarcinoma, frequently takes hold. For surgeons, the most vital question before an operation is not just whether the cancer is present, but exactly where its center lies in relation to this junction. This spatial relationship determines the entire surgical strategy: whether to remove more of the esophagus, how much of the stomach to take, and which lymph nodes to clear. Currently, doctors rely on a system called the Siewert classification to make this determination. It is a rule-based system that categorizes tumors into three types based on their distance from the junction. However, measuring this distance on standard medical scans has long been a subjective art. Radiologists look at two-dimensional slices of the body and estimate where the tumor starts and ends, a process that varies from person to person and often leads to inconsistent results, especially when tumors are large, irregular, or hidden behind other structures.

To solve this problem of human inconsistency, a team of researchers from Peking University and other institutions has developed a new digital tool called FACET. Rather than asking a computer to simply guess a tumor type, the researchers built a system that mimics the careful, step-by-step reasoning a surgeon uses. They treated the classification task not as a black-box prediction, but as a measurable geometry problem. The system takes a three-dimensional CT scan of a patient's chest and abdomen and automatically identifies the tumor and the surrounding anatomy. It then constructs a virtual path along the inner surface of the stomach and esophagus, tracing the exact route the tumor follows. By measuring the distance along this curved path from the center of the tumor to the junction, the software converts a vague visual estimate into a precise, reproducible number. This approach removes the guesswork, turning a subjective judgment into a traceable measurement that any doctor can verify.

The researchers tested this system on a large group of patients from multiple hospitals to see if it could match the accuracy of the gold standard: the physical examination of the tumor after it has been surgically removed. They divided their work into three stages. First, they taught the system using data from 70 patients. Next, they tested it on a separate group of 61 patients from the same hospital to ensure it wasn't just memorizing the first set of data. Finally, they challenged it with a completely new group of 45 patients from a different hospital to see if it would hold up in a real-world, independent setting. The results showed that the system worked well, correctly identifying the tumor type in about 73% of the cases in the final, independent group. While this is not a perfect score, it represents a significant step forward in consistency, particularly because the system achieved an area under the curve—a statistical measure of overall performance—of 0.770 in this external group, suggesting it can reliably distinguish between the different tumor types.

What makes this work distinct is how it handles the physical reality of the body. The researchers recognized that a tumor does not sit in a straight line; it follows the curves of the digestive tract. A simple straight-line measurement would be misleading. Instead, the software builds a map of the tumor's surface and calculates the distance along that surface, much like measuring the length of a winding road rather than the distance between two points on a map. To bridge the gap between the living body seen on a scan and the fixed tissue seen under a microscope after surgery, the team applied a specific calibration factor. This adjustment accounts for the fact that tissues shrink and change shape once they are removed from the body. By locking this factor at a specific value, the system ensures that the measurements taken from the scan align as closely as possible with the final pathological diagnosis.

The study also highlights the limitations of current medical imaging and the challenges that remain. The system was designed for patients who had not received prior chemotherapy or radiation, as those treatments can shrink tumors and distort the anatomy, making the measurements unreliable. Furthermore, the researchers noted that while the system is highly accurate for many cases, it still requires a human doctor to review the results, especially when the tumor boundaries are unclear or the anatomy is unusual. The tool is not a replacement for a surgeon's expertise but a powerful assistant that provides a clear, objective starting point. By making the invisible geometry of the tumor visible and measurable, this framework offers a new way to standardize preoperative planning, potentially leading to more consistent surgical outcomes and better care for patients facing this difficult disease.

Technical Summary: Functional Agent-Driven Computational Anatomy for Preoperative Siewert Classification of AEG

Problem Statement
Accurate preoperative Siewert classification of adenocarcinoma of the esophagogastric junction (AEG) is essential for determining surgical approaches, lymph node dissection extent, and prognostic stratification. However, current imaging-based workflows rely heavily on subjective estimation using axial CT or endoscopy. This subjectivity leads to poor reproducibility, inter-observer variability, and potential misclassification, particularly in complex cases involving locally advanced tumors, irregular morphologies, or indistinct tumor-organ boundaries. Furthermore, conventional endoscopy often fails to accurately delineate the longitudinal span of submucosal invasion. There is a critical need to translate the anatomical definition of Siewert classification into a standardized, computable, and traceable preoperative imaging task that moves beyond experience-dependent interpretation.

Methodology: The FACET Framework
The authors developed FACET (Functional Agent-driven Computational Anatomy EGJ Tool), a multi-agent framework designed to convert subjective spatial judgment into calibrated, traceable 3D distance measurements on contrast-enhanced CT. Unlike end-to-end "black-box" classification models, FACET decomposes the problem into a sequence of anatomically grounded, clinically measurable sub-tasks orchestrated by functional agents.

The workflow consists of four primary stages:

Perception Agent: Utilizes two anatomy-specific nnU-Net models trained on manually annotated datasets to generate candidate volumes for gastric and esophageal tumors. This stage prioritizes sensitivity to capture all relevant components, deferring strict anatomical constraints to subsequent steps.
Structure Selection and Path Modeling Agents:
- The Structure Selection Agent applies anatomical constraints to filter segmentation noise, resolve ambiguities, and select plausible components based on spatial continuity. It uses a clinician-provided tumor starting point as a spatial anchor.
- The Path Modeling Agent constructs a mucosal-surface-constrained search space along the gastric wall. It employs a 3D Breadth-First Search (BFS) algorithm starting from the anchor to generate a continuous tumor main-axis path. The tumor midpoint is defined as the point at half the geodesic path length, ensuring consistency with curved anatomical structures rather than Euclidean space.
Interface Reasoning Agent: This agent localizes the Esophagogastric Junction (EGJ) interface point using a hierarchical, evidence-driven decision strategy. It defines the EGJ anatomically as the transition between the tubular distal esophagus and the gastric cardia/fundus (distinct from the mucosal Z-line). The agent orchestrates four strategies based on available structural evidence:
- Intersection-based: Evaluates spatial overlap between gastric and esophageal tumor candidates.
- Projection-based: Selects the nearest path location to candidate overlap regions if direct linking is impossible.
- Distance-balancing: Identifies the point with the smallest difference in distance to both tumor compartments when no overlap exists.
- Lumen-density fallback: Analyzes local lumen density variations when structural evidence is insufficient.
Classification Module: Computes the geodesic distance ( $d$ ) between the tumor midpoint and the inferred EGJ interface. A calibration factor ( $k=0.7$ ) is applied ( $d' = k \times d$ ) to reconcile in vivo CT measurements with ex vivo pathological specimen dimensions. Siewert types are assigned based on calibrated distance boundaries: Type I ( $d' < -1.0$ cm), Type II ( $-1.0 \le d' \le 2.0$ cm), and Type III ( $d' > 2.0$ cm).

The entire workflow is implemented as an interactive desktop application allowing for 3D visualization, anatomical measurement, and optional clinician verification of intermediate outputs (masks, paths, interface points).

Key Contributions

Computational Anatomy Paradigm: The study reframes Siewert classification not as a direct image-to-label prediction, but as a structured computational anatomy workflow involving perception, reasoning, and rule-based decision-making.
Traceability and Interpretability: By breaking the process into modular agents, intermediate results (e.g., tumor midpoint, EGJ interface) are inspectable and correctable by clinicians, distinguishing the system from opaque deep learning models.
Anatomical Calibration: The introduction of a calibration factor ( $k=0.7$ ) addresses the systematic scale difference between in vivo CT measurements and post-resection pathological specimens, aligning the computational output with established pathological standards.
Mucosal-Surface Geodesic Modeling: The use of 3D BFS on a mucosal-constrained search space provides a more anatomically accurate representation of tumor extent compared to linear or axial measurements.

Results
The study was a retrospective multicenter diagnostic accuracy study involving 176 patients across three cohorts: development ( $n=70$ ), internal validation ( $n=61$ ), and external validation ( $n=45$ ). Histopathological Siewert classification served as the reference standard.

Development Cohort: Achieved 71.4% accuracy and an AUC of 0.721.
Internal Validation Cohort: Achieved 82.0% accuracy and an AUC of 0.927.
External Validation Cohort: Achieved 73.3% accuracy, a sensitivity of 69.7%, a specificity of 83.3%, and an AUC of 0.770. The balanced accuracy was 76.5%.

The system demonstrated moderate discrimination in the independent external cohort, with performance metrics varying based on the distribution of Siewert types across cohorts (e.g., higher Positive Predictive Value of 92.0% in the external cohort due to a higher prevalence of Type I/II tumors).

Significance and Claims
The authors claim that FACET transforms subjective imaging interpretation into a reproducible, 3D anatomical assessment. The framework is presented as a traceable decision-support tool rather than a replacement for expert assessment. Its primary significance lies in its ability to standardize the preoperative Siewert classification process, making it less dependent on individual radiologist experience and more grounded in measurable anatomical data.

The paper modestly concludes that while the independent external cohort results support the feasibility of cross-institutional application, the system currently demonstrates "moderate discrimination." The authors emphasize that further prospective validation is required, particularly for patients receiving neoadjuvant therapy (who were excluded in this study) and those with significant anatomical distortions like large hiatal hernias. The study positions FACET as a step toward reducing inter-observer variability and improving the standardization of preoperative multidisciplinary planning for AEG patients.

Technical Summary: Functional Agent-Driven Computational Anatomy for Preoperative Siewert Classification of AEG

More like this