top of page

The Anatomy of AI: Essential Terminology for Radiologists

mariomahecha0098
Nov 6
2 min read

Mario Mahecha, Santiago Aristizabal, Santiago Guzman, Roy Riascos


During a recent lab discussion, we reflected on something we frequently hear in radiology: Radiologists need to be able to critically evaluate artificial intelligence (AI).

But this raised an interesting question: What exactly are radiologists expected to evaluate? An AI system, a model, or an algorithm?

Think of transportation. Are we evaluating a car or just its engine? Each represents a different level of analysis and requires a different approach to evaluation.


The same principle applies to AI.

To understand what we are evaluating, we first need to understand how these concepts relate and how the components of an AI system work together.


From the Big Picture to the Building Blocks of AI

Artificial intelligence (AI) is the scientific and engineering field concerned with developing systems capable of performing tasks associated with intelligence, such as recognizing patterns, reasoning, and making predictions.

At a conceptual level, an AI system can be understood through three fundamental stages:


Input → Processing → Output


  • Input: The information provided to the system.

  • Processing: The computational methods that transform inputs into outputs.

  • Output: The result produced by the system.


AI is the field, an AI system is the application, and an AI model is a computational component that may power it. An algorithm, meanwhile, is a defined sequence of computational steps used to perform a task, such as processing data or training a model.


Returning to our transportation analogy, transportation engineering is the broader field, a car is a system designed to accomplish a task, and its engine is a component that enables it to function.


Importantly, evaluating an engine is not the same as evaluating an entire car. An engine may perform exceptionally well, but the car's overall performance depends on how its components work together.


Within the processing stage, different approaches can be used, including symbolic AI, which relies on explicitly defined rules, and machine learning, which uses algorithms to train models to recognize patterns in data.



Through the Rad-E-ologist's Lens

Consider an AI application designed to detect intracranial hemorrhage on head CT.

  • AI (the field): The scientific discipline behind developing intelligent imaging technologies.

  • AI system (the application): A clinical tool that analyzes head CT examinations and flags suspected hemorrhage for radiologist review.

  • Machine learning (the approach): A method that uses imaging data to train models to recognize patterns associated with hemorrhage.

  • Algorithm and AI model (the components): Algorithms define computational steps, including how a model is trained, while the trained model uses learned patterns to analyze new CT images.



The Takeaway

  • Evaluating an AI model is not the same as evaluating an AI system. A model may perform exceptionally well, but the clinical value of the entire system also depends on its reliability, workflow integration, and impact on patient care.

  • Understanding these relationships is the first step toward critically evaluating AI in radiology.


In our next article, we'll explore the different approaches used to develop AI systems, from symbolic AI to machine learning, and how they process information.

 
 
 

Comments


Rad-E-ology

  • X
  • Youtube

© 2025 Rad-E-logy. All rights reserved

bottom of page