In the 20th Century, while different forms of radiology were being developed and tested, the digital revolution was also taking hold. Radiographs no longer needed to be developed on film to be utilized. Digital images on workstations replaced films, permitting multi-planar image reconstruction [5]. Picture archiving and communication systems (PACS) have given rise to teleradiology (i.e. the transmission of digital images to locations outside of the facility where the images were obtained, allowing trained radiologists to confer their interpretations electronically and remotely.)
The increased data storage capacity of PACS systems and an emphasis on diagnostic testing has led to a boom in requests for radiology images. While this is exciting news, the human limit for interpretation has been reached, resulting in an increased risk of medical errors, as well as increased risk of fatigue and burn-out for radiologists, leading to disruption to and dilution of patient care. [5], [6]. Artificial intelligence is one promising tool to mitigate these risks.
Current applications for machine learning (ML) include pathology detection, such as that of brain hemorrhage or small lung nodules. Methods ranging from convolutional neural networks to variational autoencoders have found myriad applications in the medical image analysis field, propelling it forward at a rapid pace.
Robust artificial intelligence models can improve patient outcomes while reducing radiologists burnout. To build these models, well-defined use cases for a specific goal are needed. AI use cases should define and highlight clinical challenges potentially solvable by machine learning (ML) and automation [7], [8]. Recent focus on foundational research in AI prioritizes leveraging big data, the cloud, and ML for augmenting clinicians’ diagnoses. However, there are three major issues that arise when creating models: small, undiversified datasets, radiology’s complexity and the opaqueness of algorithm design.
Computers, software and robots are still not really “intelligent” – their abilities encapsulate a narrow focus on solving very specific questions. This limits their ability to account for uncertainty in tasks involving humans. While narrow AI currently exists for many medical applications and individually do better at tasks than trained medical professionals, the field is still far away from replacing doctors entirely. A fully automatic interventional procedure, which requires knowledge of human anatomy, real-time tracking of needles or catheters through blood vessels, and treatment of lesions is still too complicated for current machine learning techniques. For example, in interventional radiology, human-guided operation is still mainstream. Another hurdle is image de-identification and data sharing to facilitate wide availability of clinical imaging data sets. Medical algorithms require very precise training and huge datasets that are accurately labeled . It is difficult to create models that produce images suitable for human interpretation from source data leading to algorithmic bias [9].
Algorithms are often seen as “black boxes.” Comprehending the processes that underly a model’s decision-making might not be possible, which makes validation and regulation difficult. New machine learning methods for clinical imaging data, such as tailored, pre-trained model architectures and distributed machine learning methods work… but cannot be completely explained. This is especially worrisome when algorithm biases are difficult to pinpoint beyond lower accuracy thresholds for racial and gender minorities.
However, this is not an impossible task.
Today, companies like Gradient Health, Inc are bridging the gap needed to create robust medical ML models by annotating large, diverse datasets. Gradient Health addresses these issues through data sharing, annotation tools and AI integration for existing clinical workflows. Meet the best annotated medical repository on the internet. We are a team of researchers that are establishing methods to encourage data sharing for training and testing AI algorithms to promote generalizability to widespread clinical practice and minimize unintended bias. This includes automated image labeling and annotation methods as well as off the shelf annotated datasets.
Utilizing a collection of de-identified, annotated medical imaging data to foster transparent and reproducible collaborative research, Gradient assembles labeled datasets and licenses them out, thereby allowing A.I. developers to do what they do bestL build A.I., without having to worry about the hurdles involved in acquiring large volumes of training and testing data.
Meet the best annotated medical repository on the internet.
