1268. Novel Approaches for Learning Radiology Artificial Intelligence for Medical Students, Residents and Radiologists New to Computer Programming
Authors * Denotes Presenting Author
  1. Carl Sabottke *; University of Arizona
  2. Raza Mushtaq; University of Arizona
  3. Mohamed Elbanan; Penn Medicine, University of Pennsylvania Health System
  4. Bradley Spieler; Louisiana State University Health Sciences Center
  5. Khaled Elsayes; University of Texas MD Anderson Cancer Center
Artificial intelligence (AI) is becoming an integral part of clinical workflows across the globe. This has created an unmet need to educate radiologists and trainees to better understand and deploy AI in clinical spaces. Though not all radiologists are interested in learning the fundamentals of AI, those who are interested can be intimidated by the novelty of AI. Several learning tools are available across multiple platforms, but these resources are often geared toward those with a computer science background. Our goal is to make AI technology accessible to all imaging personnel including radiologists without a computer science background. We propose an innovative approach to learning the fundamentals of AI using hands-on experience with an interactive tutorial for neural network training.

Materials and Methods:
Using Jupyter notebooks, we have developed a set of tutorial modules hosted on Github which are fully runnable in the cloud-based environment of Google Colab so that radiologists and trainees without prior computer programming experience can work through developing basic neural network applications without requiring any significant hardware investment. Instead, we promote an initial experience of executing Python code via a web browser to increase accessibility. Our initial set of tutorial modules focuses on the public Combined (CT-MR) Healthy Abdominal Organ Segmentation (CHAOS) dataset and emphasizes creating U-net segmentation models using FastAI which can be applied on CT and MRI DICOM data for liver, kidney and spleen segmentation.

Our example tutorial (available at https://github.com/csabot3/liverAItutorial) focuses on training neural networks for hepatic segmentation on CT and MRI. A series of 6 Jupyter notebooks walk through the process of neural network development step by step. The first Jupyter notebook in the tutorial series reviews the basics of DICOM data handling and some basics of Python programming for beginners. This is followed by a set of notebooks that show how U-net models can be quickly trained on CT and MR data with an emphasis placed on practicality rather than granular modeling details . This is followed by a hands-on review of basic radiomics to demonstrate example applications of organ segmentations. The tutorial series then concludes with 2 additional notebooks that highlight additional modeling details and considerations with an emphasis placed on potential modeling choices that can hinder performance or lead to incorrect results.

The growing role of radiology AI in clinical workflows necessitates evolving methods of radiology education so that clinicians can understand the potential implications of radiology AI model predictions as well as the potential reasons why such predictions may be inaccurate. For radiologists and trainees with no or limited experience in computer programming and AI, we have developed hands-on exercises for neural network model development using CT and MRI data to highlight neural network model fundamentals with an emphasis on important considerations for model inaccuracies.