Photos of Pots and Toasters Help Train New Machine to Track COVID-19

Efforts to advance technology in response to the COVID-19 pandemic has led to some rather interesting breakthroughs — such as a machine learning algorithm that was trained to detect the presence of the disease by learning from everyday items.

Researchers from the Edith Cowan University (ECU) in Joondalup, Australia trained their machine learning algorithm using a dataset of over a million images of everyday items like toasters, refrigerators, pots, and carpets — objects that have little to no medical relevance to the target application. Researchers then transferred the acquired imaging knowledge of the machine learning algorithm to identify different traits associated with medical conditions that are currently diagnosed with X-ray scans.

Details of the new study are reported in the article "A novel augmented deep transfer learning for classification of COVID-19 and other thoracic diseases from X-rays," appearing in the journal Neural Computing and Applications.

ALSO READ: AI, Big Data Can't Perform Miracles in Combating COVID-19, Says Experts

Deep Transfer Learning for COVID-19 Diagnosis

The technique used by the researchers for the new machine learning algorithm is called deep transfer learning. In this method, deep learning models are trained on a different problem to be used for another application. Researchers trained their model first with everyday items seemingly unrelated to its medical imaging purposes. Their approach led to a machine learning model that has a 99.24 percent success rate in finding COVID-19 from chest X-rays.

Aside from its potential in the medical imaging field, the study also helps address one of the largest challenges today in the field of machine vision, image recognition, and machine learning in general. This refers to the new model's ability to work around the limited medical imaging data available. To be accurate in its tasks, algorithms generally require large datasets to be able to recognize predetermined target characteristics. By using deep transfer learning, researchers were able to pre-train the machine learning model before being used for its intended application.

"Our technique has the capacity to not only detect COVID-19 in chest X-rays, but also other chest diseases such as pneumonia. We have tested it on 10 different chest diseases, achieving highly accurate results," shares Dr. Shams Islam, one of the authors and a researcher from the ECU School of Science, in a news release from the university. He noted that normally, it is "difficult for AI based methods to perform detection of chest diseases" like COVID-19 accurately due to the need for large amounts of training data, which is used by the algorithm to understand the signatures of each medical condition.

He adds that while the novel method remains unlikely to replace the rapid COVID-19 tests today, it still has important implications in other medical imaging needs.

A Machine Learning Shortcut

Fouzia Atlaf, lead author and Ph.D. candidate from ECU, explains that the key to greatly reduce the training time needed by the deep transfer learning algorithm was to pretrain it with the large ImageNet database. The huge compilation of images in ImageNet are organized according to the WordNet hierarchy, each being hand-annotated to indicate in text the contents of the image. It was developed specifically to help train AI algorithms. The hand-annotation is similar to the human classification and notes made by medical professionals in chest X-rays.

"The difference is the images in the database are of regular household items which can be classified by people without medical expertise," Atlaf explained.