ARTIFICIAL INTELLIGENCE
( continued from page 9 )
Advanced Artificial Intelligence ( AAI ): This seems to be interchangeable with or more recently replacing AI with respect to diagnostic imaging .
Machine Learning ( ML ): A subset of AI that uses statistical methods to enable machines to improve with experience .
Deep Learning ( DL ): A subset of ML that makes the computation of multi-layered neural networks feasible .
DL models learn discriminatory features that best predict outcomes i . e ., recognize patterns in the data . They can detect specific features ( present or absent ), classify ( benign vs . malignant ), localize , measure or quantify , predict risk .
There are currently five use designations for AI algorithms : 4 1 ) Computer Aided Detection ( CADe ) 2 ) Computer Aided Diagnosis ( CADx ) 3 ) Computer Aided Detection and Diagnosis 4 ) Computer Aided Triage and Notification ( CADt ) 5 ) Radiologic Acquisition and Optimization Guidance
The creation of any AI algorithm that is going to be successful in diagnostic imaging and therefore useful in clinical practice requires the input of large amounts of data . The information contained in the data must be accurate and precise . Imaging findings must be correctly identified in relation to diagnosis . Information taken from the electronic medical record must be current and without error . Data scientists and software programmers are key to the production , but radiologists should be involved in evaluation of the clinical application and validation of the structured use cases . The algorithms should have a narrow use application , for example , detecting pneumothorax on chest X-rays 5 or intracranial aneurysms on head CTA ( computed tomographic angiography ). 6
Large research institutions with their own data scientists and IT departments have been creating algorithms for their own institutional use for a number of years . Radiologists are an essential component of these teams . Small practices and hospital systems will rely on commercially available AI models . In 2017 , in order to bring these areas of interest together , the American College of Radiology ( ACR ) created the Data Science Institute ( DSI ) R www . acrdsi . org .
One of the main functions of the DSI is to allow external validation of algorithms . Scientific institutions , hospital communities , practice groups and commercial software vendors can submit algorithms for test use . The broader the base of clinical input , in particular patient demographics , and the higher the incidence of disease in any data set , the likelier will be the final success of the algorithm . The primary imaging goal will always include optimum levels of sensitivity and specificity .
Since all data is patient generated , any sharing outside institutional boundaries becomes problematic with respect to HIPAA compliance . All information must be anonymized and maintained in a secure environment . DSI has created a secure cloud where data can be held and accessed as needed . I believe it is true to say that clinical radiologists are somewhat ambivalent about the role of AI in their practices . DSI maintains a catalog of use sets indicating what algorithms are commercially available ; what algorithms are under construction ; what algorithms have been externally validated . The goal is to enable radiologists to understand and be proactive in their use of AI . Recognizing that residents in training are open to the use of AI in their future practices , DSI has a training program hosted on its AI- LAB TM platform that incentivizes residents to create their own algorithms . Starting in November 2021 , the program was expanded and is now offered to practicing radiologists who want to be a part of the future of AI . Software programmers can produce algorithms , but they will serve no useful purpose unless the importance of the clinical input is recognized in the effectiveness of the end product : more accurate diagnosis and / or improved workload management .
In addition , radiologists will not be receptive to AI if the systems do not have interoperability with the software programs that radiologists are using in their everyday interpretive functions . The time is over for radiologists to act as human toggles between financially competing imaging programs .
“ Opinions on the future of AI in diagnostic imaging range from the apocalyptic claim that AI will make all radiologists extinct to the delusional assertion that computers will always merely assist . These views are both mistaken .” 7 It is up to radiologists to decide how far the arc will bend toward the former .
References :
1
ArtificiaI Intelligence and the Practice of Radiology ; Schier R ., MD , JACR online : Published May 11 , 2018 DOI : https :// doi . org / 10.1016 / j . jacr . 2018.03.04
2
Applications of Pixel-Based AI ; Gichoya J . W ., MBChB ., MS . 2020 Imaging Informatics Summit
3
Creating Artificial Intelligence ; Dryon K . DO ., PhD ., FACR ., FSIIM . 2020 Imaging Informatics Summit .
4
Evaluating AI for use in your Clinical Practice : Allen B Jr ., MD , FACR . Informatics e-Learning Hub . ACR AI-LAB TM
5
AI-assisted Pneumothorax Detection on Chest X-rays ; Hong W . MD , et al . Radiology online :
Published Jan . 25,2022 DOI : https :// doi . org / 10.1148 / radiol . 211706
6
Artificial Intelligence Assistance Improves the Accuracy and Efficiency of Intracranial Aneurysm Detection with CT Angiography ; Wei Xin et al ., European Journal of Radiology : Published January 20 , 2022
DOI : https :// doi . org / 10.1016 / j . ejrad . 2022.110169
7
Schier R ., MD , in Artificial Intelligence and the Practice of Radiology ; JACR online : See ( 1 ) above
Dr . Amin is a retired diagnostic radiologist .
10 LOUISVILLE MEDICINE