Unsupervised Chest X-ray Opacity Classification using Minimal Deep Features

Data privacy has been a concern in medical imaging research. One important step to minimize the sharing of patient's information is by limiting the use of original images in the workflow. This research aimed to use minimal deep learning features in detecting anomaly in chest X-ray (CXR) images. A total of 3,504 CXRs were processed using a pre-trained deep learning convolutional neural network to output ten discriminatory features which were then used in the k-mean algorithm to find underlying similarities between the features for further clustering. Two clusters were set to distinguish between Opacity and Normal CXRs with the accuracy, sensitivity, specificity, and positive predictive value of 80.9%, 86.6%, 71.5% and 83.1%, respectively. With only ten features required to build the unsupervised model, this would pave the way for future federated learning research where actual CXRs can remain distributed over multiple centers without sacrificing the anonymity of the patients.

Automated summary

This study uses minimal deep learning features to detect anomalies in chest X-ray images and identifies opacity and normal images through clustering algorithm. The results show that an unsupervised model can be built with only ten features, paving the way for future federated learning research.