Deep Learning and Transfer Learning used to detect ARDS on Chest X-rays

In this article by Sjoding et al. in The Lancet Digital Health the authors

sought to train a deep convolutional neural network (CNN) to detect ARDS findings on chest radiographs.

They were able to produce concistent results:

a CNN could detect ARDS with an area under the receiver operator characteristics curve (AUROC) of 0·92 (95% CI 0·89–0·94). […] In an external cohort of 958 chest radiographs from 431 patients with sepsis, the AUROC was 0·88 (95% CI 0·85–0·91).

They were presented with a requiring problem in AI assisted digital health:

Generating large datasets for many medical problems can be challenging because clinical data might not be annotated for the finding of interest during routine care.

They were able to use Transfer Learning to accomodate this challenge:

Transfer learning is a machine-learning approach where knowledge gained from one problem can be used to help solve related problems. […] When training a CNN to detect ARDS, we hypothesised that if the network could first learn to extract general features from chest radiographs by pretraining the CNN to identify other common findings on chest radiographic studies, it might be able to borrow many of these features, reducing the number of annotated images necessary to train the network to detect findings of ARDS. […] We used transfer learning by first pretraining the network on 595 506 radiographs from two centres labelled for common descriptive chest findings (eg, opacity, effusion), but not ARDS. We then trained the network on 8073 radiographs annotated for ARDS. We tested the resulting network on an internal and external test set to evaluate its generalisation performance.

In summary, these results show the power of deep learning models, which can be trained to accurately identify chest radiographs consistent with ARDS. Further research is needed to evaluate how the use of these algorithms could support real-time identification of ARDS patients to ensure fidelity with evidence-based care or to support ongoing ARDS research.

The researchers also were also able to visualise the CNN activations on the chest X-rays. In a clinical setting this can help physicians quickly assess the accuracy of the algorithm.

Their study did have some limitations. The researchers state that ARDS is a syndrome and is not defined by a Gold standard, making uniform diagnosis impossible.

I also found basis for possible race bias. There’s a major race inconsistency between the training and external datasets (9% vs 30% blacks).