Introduction
Gastrointestinal (GI) bleeds pose significant challenges to clinicians due to their potential severity and intermittent nature. Traditional diagnostic methods involve visual analysis of computed tomography (CT) scans by radiologists, which can be time-consuming and prone to subjective interpretation. To address these challenges, artificial intelligence’s (AI) unsupervised deep learning models offer a promising solution by automating the process of bleed detection and potentially improving diagnostic accuracy.
Methods
Variational Autoencoder (VAE), an unsupervised AI model for GI bleed detection in abdominal CT scans is utilized. Its architecture is composed of reconstruction and predicting geometric transformations using fully connected 3 hidden layers for both encoder and decoder in addition to transposed convolutions. Training with online scraped dataset of total 276,351 abdominal “normal” CT scans was performed to predict geometry. VAE constructed “normal”, and “abnormal” scans dissimilarity vector to find bleed. Adaptive learning rates during neural network training ranging from 0.00001 to 0.005 were used to facilitate steady convergence.
Results
Preliminary results undertaken on tests over 658 “abnormal” and 7,629 “normal” images suggest that our model achieved ~92% sensitivity and ~90% specificity in classifying bleed versus non-bleed cases. An Area under the ROC Curve of ~0.907 was found. The model identified the spatial coverage of the bleed in the abnormal scan images with ~89% in-pixel precision when compared to traditional manual methods. Additionally, 897 artificial bleed cases were generated by changing a single pixel contrast at random locations on normal CT scan images. With this, model’s true positive rate was reduced to ~0.61.
Discussion
Our unsupervised images reconstruction approach identified its own pattern of abnormalities in contrast to most other CT scan-reading-AI models that utilized annotated or labelled data for training. This way we overcame the bias and limitations of annotated abnormal cases by humans. Our motivation is to be able to detect bleeds in patients with definite clinical bleeding signs but negative CT scan reads. With artificial contrast testing, we recognize that AI model requires exploration into architectural adjustments of local reconstruction. We foresee challenges such as difficulty in detecting smaller anomalies and accurate localization in gastrointestinal tract. There is also a need for consideration for data quality, heterogeneity and variations in imaging practices. Ethical considerations regarding bias and fairness necessitate vigilant scrutiny for equitable performance across demographic spectra. The next step in this research is a validation trial of a real-world collection of CT scans compared to radiology reads and clinical outcomes to more firmly establish its accuracy.

Shefali Amin

Appearances