Development of CNN models for the enteral feeding tube positioning assessment on a small scale data set

Gongbo Liang, Halemane Ganesh, Dylan Steffe, Liangliang Liu, Nathan Jacobs, Jie Zhang

Introduction

Ensuring the correct placement of an enteral feeding tube (EFT) is a critical safety procedure in clinical settings. Misplacement—such as a tube entering the lungs instead of the stomach—can lead to severe, even fatal, complications. Currently, the gold standard for confirmation is the manual review of chest radiographs by clinicians. However, this process is resource-intensive and can lead to delays in patient care.

While deep learning, specifically Convolutional Neural Networks (CNNs), has shown immense potential in medical imaging, these models typically require vast amounts of manually labeled data. In the context of specialized tasks like EFT positioning, obtaining thousands of expert-labeled radiographs is often prohibitively expensive and time-consuming. This research addresses the "small data" challenge by proposing a weakly supervised pre-training strategy that leverages existing clinical data to build robust models with minimal manual labeling.

Methodology: Weakly Supervised Pre-training

The core innovation of this study is a pre-training approach that bridges the "domain gap" between general-purpose image datasets (like ImageNet) and medical radiographs. Instead of training on photos of everyday objects, the researchers developed a dual-branch network designed to learn from the inherent relationship between medical images and their corresponding textual reports.

Illustration of the weakly supervised pre-training approach

As shown in the architecture above, the system consists of a text branch (using a BERT encoder) and an image branch (using a ResNet-18). By processing over 360,000 image-report pairs from the MIMIC-CXR dataset, the model learns to recognize complex anatomical structures and medical features through contrastive learning—predicting whether a report and image are a "natural match" without requiring explicit human-drawn labels.

Fine-Tuning for EFT Assessment

After the initial pre-training phase, the image feature extractor is repurposed for the specific task of feeding tube assessment. The researchers appended a classification head to the pre-trained ResNet-18 and fine-tuned it using a small, specialized dataset of 175 radiographs (112 showing satisfactory placement and 63 showing unsatisfactory placement).

The enteral feeding tube positioning assessment network

This fine-tuning process allows the model to specialize its learned medical knowledge toward identifying the presence and correct positioning of the feeding tube tip, even when trained on a very limited number of labeled examples.

Results and Performance

The proposed weakly supervised approach significantly outperformed baseline models across all key performance indicators:

  • Discriminative Power: The model achieved an Area Under the Curve (AUC) of 0.76. In comparison, a model trained from scratch (no pre-training) achieved only 0.56, while a standard model pre-trained on ImageNet reached 0.67.
  • Classification Accuracy: The system reached an accuracy of 78%, demonstrating that pre-training on domain-specific medical data is far more effective than using general-purpose datasets when labeled data is scarce.
  • Model Calibration: Beyond simple accuracy, the researchers measured the Expected Calibration Error (ECE). The proposed model achieved the lowest ECE (0.09), meaning its predicted confidence levels closely align with its actual probability of being correct—a vital trait for clinical decision-support tools.

Discussion and Future Directions

The study demonstrates that "annotation-efficient" modeling is possible in the medical domain by leveraging the "free" labels found in existing radiology reports. By using contrastive learning to align text and images, the model develops a sophisticated understanding of chest anatomy that generalizes well to specific tasks like EFT assessment.

However, limitations remain. The evaluation was conducted on a relatively small dataset of 175 images from a single institution. Furthermore, the model's performance can be degraded by visual noise or the presence of other medical devices. Future work will focus on expanding the test set across multiple clinical sites and implementing "boosting" techniques to improve robustness in complex, high-noise environments.

Summary updated

Abstract

Enteral nutrition through feeding tubes serves as the primary method of nutritional supplementation for patients unable to feed themselves. Plain radiographs are routinely used to confirm the position of the Nasoenteric feeding tubes the following insertion and before the commencement of tube feeds. Convolutional neural networks (CNNs) have shown encouraging results in assisting the tube positioning assessment. However, robust CNNs are often trained using large amounts of manually annotated data, which challenges applying CNNs on enteral feeding tube positioning assessment.
BibTeX
@article{liang2022development,
  author = {Gongbo Liang and Halemane Ganesh and Dylan Steffe and Liangliang Liu and Nathan Jacobs and Jie Zhang},
  title = {{Development of CNN models for the enteral feeding tube positioning assessment on a small scale data set}},
  journal = {BMC Medical Imaging},
  year = {2022},
  doi = {10.1186/s12880-022-00766-w},
}
Figure