While the images of artificial intelligence popularized in modern film and literature are still far from a reality today, subsets of the field like machine learning and deep learning continue to make significant advancements. Machine learning uses statistical models, patterns, and inference to teach computer systems to conduct tasks without requiring explicit instructions. By drawing from a set of training data, computer systems are then able to build up a mathematical model to make decisions and predictions without relying on strict programming. One of the most successful and pervasive achievements of this form of machine learning, or more specifically deep learning and deep neural networks, is automatic speech recognition, which now goes by many names, such as Alexa, Siri, or Cortana.
In addition to speech recognition, one of the areas that machine learning excels in is image analysis, which has major implications for medical imaging and diagnosis. Current imaging technologies, like CT scans, MRI scans, and X-rays can be time consuming in comparison to the image analysis conducting by machine learning algorithms. More efficient and accurate radiology tools aided by AI could also reduce the need for tissue samples, allowing for remote diagnosis for patients in rural areas.
For example, researchers at the Aichi Cancer Center Hospital and the Yokohama City University School of Medicine in Japan published their study on the ability of AI to effectively differentiate between malignant and benign cystic lesions. The team tested the AI’s deep learning by retroactively inputting a data set of 85 patients who underwent cyst fluid or surgical specimen analysis to diagnose malignancy and the AI performed significantly better in diagnosis than traditional methods. In a 2018 study of deep learning with chest radiograph diagnosis, researchers at Stanford University used an algorithm, called CheXNeXt, to interpret 420 images of pathologies like pneumonia, pleural effusion, pulmonary masses, and nodules. In comparison to a group of 9 radiologists who took 240 minutes to review the same images, the model analyzed the images in 1.5 minutes and with comparable accuracy.
One of the primary obstacles for deep learning in medicine is the huge amount of data that is required in order to build a model that will make accurate decisions and predictions, and this resource is limited by privacy laws for the use and distribution of patient data. There may also be issues with the format or quality of the collected data that would require more human effort to prepare the training sets for machine learning analysis. Furthermore, it is important to be cognizant of biases that may exist in data sets that do not include a diverse range of demographics.
Currently, most machine learning research remains small in scale and limited in scope because of these obstacles, but the implications for the world of image analysis and diagnosis in medicine are enormous. In the coming decades, machine learning may overwhelmingly replace the interpretive work of radiologists with more efficient and accurate diagnoses.