TL;DR
This paper introduces a deep learning system that automatically segments the trachea and endotracheal tube (ETT) from chest X-rays and then classifies the ETT's position into four categories: absent, shallow, proper, or deep. This approach shows promise in improving the accuracy and efficiency of ETT placement assessment, a critical step in patient care.
What Problem Does This Paper Address?
Proper placement of an endotracheal tube (ETT) is crucial for effective airway management in critically ill patients. An ETT that is too shallow can lead to accidental extubation or aspiration, while one that is too deep can cause bronchial intubation, lung hyperinflation, and barotrauma. Chest radiography is a common tool for confirming ETT placement, but visually assessing the distance from the carina to the ETT tip can be challenging, especially under pressure or with poor image quality. Existing automated methods have shown success in detecting the ETT's presence but struggle with precise positional classification. This paper aims to bridge this gap by developing an end-to-end deep learning system for both automatic segmentation and accurate positional classification of the ETT.
Methods
The researchers propose a two-stage deep learning approach. First, they employ a Mask R-CNN model, pre-trained on ImageNet and then fine-tuned on their dataset, to automatically segment the trachea and the ETT from chest X-ray images. This segmentation step is crucial for providing precise anatomical context. Second, they use an EfficientNet B0 model for the multi-label classification of the ETT's position. This classification model takes a 3-channel input: the original grayscale X-ray, the segmented trachea mask, and the segmented ETT mask. These inputs are then cropped using bounding boxes predicted by Mask R-CNN and resized for the EfficientNet. This combined segmentation-classification strategy is designed to enhance the accuracy of positional assessment.
Datasets
The study utilized a dataset of 1985 chest anteroposterior (AP) radiographs collected from patients who underwent ETT intubation at a tertiary hospital between January 2018 and March 2021. The dataset was divided into:
- Training and Validation Set: 596 images were manually segmented for the trachea and ETT. These were further categorized into four ETT position labels: absence (n=298), shallow (n=97), proper (n=100), and deep (n=101).
- Test Set: 1389 non-segmented images were used for evaluating the performance of the classification model after it was trained on segmented data.
Key Insights and Ideas
- Segmentation is Key: The study highlights that accurate segmentation of the trachea and ETT is a critical precursor to reliable positional classification. By providing these masks as input to the classifier, the model gains a more robust understanding of the anatomical context.
- Two-Stage Approach Synergy: Combining a powerful segmentation network (Mask R-CNN) with a state-of-the-art classification network (EfficientNet) creates a synergistic effect, leading to improved performance over using either component in isolation.
- Multi-Label Classification for Nuance: Moving beyond a simple "proper" vs. "improper" classification, the four-label system (absent, shallow, proper, deep) provides a more granular and clinically relevant assessment of ETT placement.
- Real-World Applicability: The use of plain chest radiographs, a readily available imaging modality, makes this approach highly practical for clinical settings.
Key Results
- Segmentation Performance: In five-fold validation on segmented images, the model achieved Dice coefficients of 0.841 (0.063) for the trachea and 0.893 (0.078) for the ETT, indicating good overlap between predicted and ground truth masks.
- Classification Performance (Segmented Data): During five-fold validation with segmented images, the four-label classification accuracy reached 0.945 (0.017).
- Classification Performance (Non-Segmented Data): On the test set of 1389 non-segmented images, the model achieved overall values of 0.922 for accuracy, 0.843 for precision, 0.843 for sensitivity, 0.922 for specificity, and 0.843 for F1-score.
- Label-Specific Performance: While overall performance was strong, accuracy was highest for the "absence" label. The "proper" position showed the lowest accuracy (0.853) and F1-score (0.325) in the non-segmented test set, though sensitivity was higher for shallow and deep positions than proper.
Limitations and Future Work
The authors acknowledge several limitations:
- Single-Center Data: The model was trained on data from a single center, which might limit its generalizability to other hospital environments.
- Definition of "Proper" Position: The specific distance range (30-70 mm from the carina) used to define "proper" position might vary across different institutions.
- Esophageal Intubation: The dataset did not include images of esophageal intubation, meaning the model cannot currently detect this critical misplacement.
- Influence of Lung Conditions: The study did not investigate how various lung pathologies (e.g., pleural effusion, pneumothorax) might affect the model's performance.
- Clinical Outcome Comparison: A comparison with physician performance regarding clinical outcomes, diagnostic time, and required equipment was not performed.
Future work could involve expanding the dataset to include diverse patient populations and pathologies, validating the model in multi-center studies, and directly comparing its performance and impact on clinical decision-making against human experts.
Summary updated