Self-Supervised Deep Learning in the Assessment of Enteral Feeding Tube Positioning Based On A Small Dataset
G Liang1*, D Steffey2, H Ganesh3, J Zhang4, (1) Eastern Kentucky University, ,,(2) University Of Kentucky, ,,(3) University Of Kentucky, ,,(4) University of Kentucky, Lexington, KY
Presentations
SU-IePD-TRACK 1-3 (Sunday, 7/25/2021) 5:30 PM - 6:00 PM [Eastern Time (GMT-4)]
Purpose: Enteral nutrition (EN) through feeding tubes serves as the primary method of nutritional supplementation for critically ill patients unable to feed themselves. The position of the Nasoenteric feeding tubes is routinely confirmed by plain radiographs after insertion and before the commencement of tube feeds. This work uses a self-supervised training strategy to improve the DCNN performance on an extremely small dataset.
Methods: A self-supervised pre-training strategy called Comparing to Learn (C2L) was used as the backbone of our EN position DCNN model, which included a batch-level mix-up method for data augmentation and a novel teacher-student network for feature learning. The plain abdominal radiographs of 100 patients were retrospectively retrieved and used in this study. The images were resized to 256x256 and equally split into five folds for five-fold cross-validation. Different windows of pixel values (0-4095) were tested. Nine DCNN architecture were investigated. In total, over 100 models were trained and evaluated. Each model was trained five times. The accuracy and area under the receiver operating characteristic (ROC) curve (AUC) on the test data were used to assess the models. The ground-truth classification for enteric feeding tube positions was performed by a board-certified abdominal radiologist.
Results: The preliminary results show that, by including a self-supervised pre-training strategy, DL provides an encouraging solution in the binary classification of the appropriate tube positioning vs. wrong positioning based on the extremely small dataset with an accuracy of 0.76 (95% CI 0.75-0.77) and AUC of 0.73 (95% CI 0.72-0.74), which is also an over 7% improvement of accuracy and over 5% improvement of AUC compared with the ImageNet pre-trained weights. Windowing setting (0-2750) helps improve the performance of DL models.
Conclusion: Self-supervised DL based on a small dataset demonstrates promise in the assessment of appropriate vs inappropriate EN feeding tube placements.
Keywords
Radiography, Image Analysis, CAD
Taxonomy
IM/TH- Image Analysis (Single Modality or Multi-Modality): Machine learning
Contact Email