The increased availability of highly effective tools and image data sets has made AI use for medical imaging research more accessible than ever before.In radiology, artificial intelligence (AI) is the engineering of computerized systems that can perform tasks typically done with human intelligent behavior such as acquiring, reconstructing, analyzing, and/or interpreting medical images. Machine learning (ML) algorithms are a subset of artificial intelligence methods that computerized systems use to solve problems by recognizing patterns in the data. AI has been developed to improve radiology workflows and assist radiologists with tasks such as lesion detection and quantification of medical images. Those exploring artificial intelligence and machine learning in medical imaging research require software tools to assist in algorithm development and image data to train and test algorithms.
Many tools and resources exist that can help develop AI algorithms for medical imaging such as:
Figma, which allows multiple team members to create wireframes and mockups of proposed designs;
TensorFlow, which provides a collection of workflows to develop and train models using Python, JavaScript, or Swift;
Tribuo, which provides a Java Machine learning library to code algorithms;
ML.NET, a software machine learning library for the C# and F# programming languages;
PyTorch, a machine learning framework with a Python interface and C++ interface; and
Keras, a library that provides a Python interface for artificial neural networks. These tools can assist developers to build and then later deploy their machine learning applications based on their preferences and expertise.
Most developing AI algorithms cannot directly access a picture archiving and communication system (PACS) environment. Traditionally, there has been a perception that a lack of adequate medical image data exists due to small sample sizes and a lack of geographic diversity. However, some large datasets now have high quality images and annotations. These datasets have gone through DICOM de-identification to meet the U.S. Health Insurance Portability and Accountability Act (HIPAA) requirements. These datasets also include image labels, which are annotations performed by radiologists. Some AI algorithms need annotations, which serve as ground truth, for medical image classification based on a supervised learning approach. The table below offers a list of accessible medical imaging datasets. These datasets include those available from Kaggle, the Cancer Imaging Archive (TCIA), the National Institutes of Health (NIH), and Stanford University’s Center for Artificial Intelligence in Medicine.
