TL;DR
This paper introduces ENDOANGEL-GEV, an interpretable AI system that accurately detects and quantifies risk factors for gastroesophageal variceal bleeding (GEV) from endoscopic images. By providing visualized, objective, and quantitative risk assessments, ENDOANGEL-GEV aims to improve GEV diagnosis, personalize treatment, and ultimately enhance patient outcomes.
What Problem Does This Paper Address?
Gastroesophageal varices (GEV) are a serious complication of cirrhosis, with bleeding being a life-threatening event. Endoscopy is the gold standard for diagnosing GEV and assessing bleeding risk, guiding prophylactic treatment. However, the interpretation of endoscopic findings, such as varix grade and red color signs (RC), is highly subjective and prone to inter-observer variability among endoscopists. This inconsistency can lead to misdiagnosis, resulting in either unnecessary treatment for low-risk patients or missed opportunities for high-risk patients, both of which have significant clinical and economic consequences. A quantitative and objective method for assessing GEV risk factors is urgently needed.
Methods
The researchers developed an AI system called ENDOANGEL-GEV, which comprises several deep learning models. The core of the system includes:
- Segmentation Models: Models for segmenting esophageal varices (EV) and gastric varices (GV), as well as red color signs (RC) on these varices.
- Classification Models: Models for classifying the grade (size) of varices and the classification of RC.
- Supportive Models: These models are crucial for ensuring data quality by filtering out unqualified images (e.g., blurred, poorly prepared, or containing biopsy forceps) and standardizing the input for the main models.
- Interpretable Interface: A key innovation is the system's interpretability. It visualizes the risk factors directly on the endoscopic images, showing probability maps and delineating identified features. This allows end-users to understand why the system makes a particular assessment.
The system utilizes fully convolutional networks (FCNs) for segmentation and deep convolutional neural networks (DCNNs) for classification. For RC classification, a density-based spatial clustering of applications with noise (DBSCAN) algorithm is employed to quantify the distribution and density of red color signs.
Datasets
The study utilized three datasets:
- Dataset 1: A large dataset comprising 6034 images from 1156 cirrhotic patients across three hospitals, used for training and validation. This dataset was further split into training, validation, and testing subsets.
- Dataset 2: A multicenter validation dataset consisting of 11009 images from 141 videos, collected from three different hospitals. This dataset was used to assess the system's generalizability across various imaging equipment and settings.
- Dataset 3: A prospective study dataset involving 161 cirrhotic patients from Renmin Hospital of Wuhan University. This dataset was used to compare the performance of ENDOANGEL-GEV against experienced endoscopists in a real-world clinical setting.
Key Insights and Ideas
- Bridging the Subjectivity Gap: The most significant insight is the successful development of an AI system that can objectively quantify GEV risk factors, directly addressing the inherent subjectivity of human interpretation in endoscopy.
- Interpretable AI for Clinical Trust: The emphasis on interpretability, by visualizing the AI's decision-making process on the images, is crucial for building trust among clinicians and facilitating adoption. This moves beyond "black box" AI.
- Comprehensive Risk Assessment: The system doesn't just detect varices; it quantifies their grade, the presence and characteristics of red color signs, and their distribution, providing a more nuanced risk assessment than traditional methods.
- Generalizability Across Diverse Data: The validation on a multicenter dataset with varying imaging equipment demonstrates the robustness and potential for broad applicability of ENDOANGEL-GEV.
- Potential for Improved Patient Care: By accurately identifying high-risk patients and reducing misclassifications, the system has the potential to significantly improve treatment efficacy and reduce healthcare costs.
Key Results
ENDOANGEL-GEV demonstrated strong performance across all datasets:
- Dataset 1 (Testing): Achieved high accuracy in segmenting esophageal varices (mIoU 0.8087) and gastric varices (mIoU 0.8141). It also showed excellent performance in RC detection and classification, and varix grade classification (e.g., 93.00% accuracy for grade classification).
- Dataset 2 (Validation): Maintained high accuracy, with EV segmentation sensitivity of 90.79% and mIoU of 0.8890. RC detection and classification also performed well.
- Dataset 3 (Prospective Study):
- EV Detection: Comparable sensitivity to endoscopists (100.00% vs. 99.23%).
- RC Classification: Significantly higher accuracy than endoscopists (94.62% vs. 66.92%, p < 0.001).
- Grade Classification: Significantly higher accuracy than endoscopists (94.57% vs. 75.97%, p < 0.001).
- Risk Stratification (EV): ENDOANGEL-GEV significantly outperformed endoscopists (97.69% vs. 85.38%, p < 0.001).
- The system could potentially help 12.31% more patients receive the right intervention.
Limitations and Future Work
The authors acknowledge several limitations:
- Dataset 3 Patient Demographics: A significant portion of patients in the prospective study had advanced liver failure (Child-Pugh C), which might limit generalizability to populations with less severe cirrhosis.
- Single-Arm Study Design: While comparisons were made, the prospective study was single-arm, not a randomized controlled trial.
- GV Location Classification: The system did not classify GV by location, as endoscopic location is not the gold standard for determining supplying and drainage vessels. Contrast-enhanced CT is still preferred for this.
Future work could involve:
- Expanding the system's capabilities to include GV location classification.
- Conducting larger, randomized controlled trials to further validate the system's clinical impact.
- Investigating the system's performance across a wider range of patient demographics and disease severities.
- Further refining the interpretability features to provide even more actionable insights for clinicians.
Summary updated