Rules to Reasoning: Evolution of AI & its Scale

+

⚗

IRHAI Institute For Responsible
Healthcare AI

Author: Dr. Sharad Maheshwari MD |
imagingsimplified@gmail.com

🧠 Rules to Reasoning

Evolution Foundations Perceptrons & MLPs Epochs Clinical Matchmaker Clinical Risks Governance

The Architecture of Radiology AI

The Evolution of AI is the Evolution of Scale

Humans did not invent matches because stones stopped producing fire. Humans invented matches because fire needed to scale. Humans did not invent lighters because matches stopped working. Humans invented lighters because fire needed to become more reliable, portable, and universally accessible.

The history of AI follows the same pattern. Every generation emerged because the previous one hit a bottleneck:

Symbolic AI → Could not scale Knowledge

Machine Learning → Could not scale Data/Features

CNNs → Could not scale to Language

Transformers → Could not scale to Many Tasks

Foundation Models → Could not scale to Workflows

Agentic AI → The next bottleneck is Governability

Foundations & The Epistemological Divide

Before algorithms, we must understand how a machine approaches knowledge and behavior.

⚞

Symbolic AI

Knowledge as Rules

The earliest CAD systems encoded deterministic logic trees based on expert heuristics. Knowledge is explicitly programmed.

• Input: Human-encoded rules
• Example: IF nodule > 8mm THEN suspicious
• Limitation: Brittle; rule explosion occurs quickly.

📊

Statistical AI

Knowledge as Pattern

Knowledge is not encoded but discovered via statistical relationships. The machine learns from vast datasets.

• Input: Data and patterns
• Example: Deep Learning, Foundation Models
• Limitation: The system does not "know" medicine.

⚙

Clarifying "Deterministic AI"

A common pedagogical error is confusing Expert Systems with Deterministic Systems. Determinism is not an AI type; it is a behavioral property. It means: Same Input → Same Output, Every Time.

System Type	Symbolic?	Deterministic?	Example
Clinical Score (MELD)	Yes	Yes	Input → Formula → Score
Expert System (MYCIN)	Yes	Usually Yes	KB + Rules + Inference Engine
Frozen Random Forest	No (Statistical)	Yes	Learned patterns, but frozen execution.
Modern LLM (GPT-4)	No (Statistical)	No (Probabilistic)	Generative probability distributions.

The Neural Blueprint

Perceptrons & MLPs

How machines moved from static rules to learning representations.

1. The Perceptron (The First Learning Machine)

Instead of writing rules, we feed inputs into a node. The machine learns Weights (how important a feature is) and a Bias (a baseline tendency). Use the sliders below to see how a single non-linear perceptron calculates clinical risk.

⚙

Interactive Tool: Artificial Neuron Simulator

Feature 1: Nodule Size 0.5

Weight (Importance) 0.8

Feature 2: Spiculation 0.2

Weight (Importance) 0.6

Bias (Clinical Threshold) -0.5

Math: (F1×W1) + (F2×W2) + Bias = Z

Z = 0.02

Low Risk Activation (Sigmoid) High Risk

Prediction: 50%

Indeterminate

2. Multi-Layer Perceptron (Solving Non-Linearity)

Biology is not linear. To solve complex patterns, scientists stacked perceptrons into Hidden Layers. This enables Representation Learning, where the network automatically extracts increasingly abstract and complex features from raw data.

Input Layer

→

Hidden 1

Edges

→

Hidden 2

Shapes

→

Output

Diagnosis

The 5 Epochs of Radiology AI

Select an epoch below to explore its clinical role and primary limitations.

Clinical Synthesis

Clinical Case Matchmaker

Which algorithm is required for a specific task? Select a clinical scenario to see the architectural rationale.

Recommendation

Architectural Selection

Select a clinical challenge above.

Formula: ...

Clinical Realities & Risks

A critical appraisal of specific architectures currently impacting the field.

Convolutional Neural Networks

CNNs revolutionised image processing by automatically learning features (edges → textures → shapes → disease). However, they are highly susceptible to Domain Shift.

The Zech et al. (2018) Evidence [10]

A model trained to detect pneumonia degraded substantially when applied to data from a hospital not represented in training. It learned statistical shortcuts (institutional metadata artifacts) rather than purely diagnostic features.

Required reading for evaluating single-institution AUC claims.

⚠ Devil's Advocate

Domain shift critique shouldn't dismiss CNNs wholesale. Multi-site studies show maintained performance when training data is demographically diverse. The problem is the validation standard, not the architecture.

Foundation Models & The Validation Gap

Defined by large-scale pretraining on diverse data followed by task-specific fine-tuning. Despite impressive technical benchmarks [3], clinical integration is constrained.

• Evaluated mostly on curated benchmark datasets.
• Benchmarks do not reflect real-world distributional heterogeneity, temporal variation, and edge-case density [5].
• Continuous evaluation and recalibration are essential governance requirements [4].

⚠ Devil's Advocate

The "we need more studies" argument can calcify into an unfalsifiable standard. Are we designing validation frameworks proportionate to clinical risk, or reflexively applying demanding standards to technology that threatens professional roles?

VLMs and the Reporting Frontier

Automated report generation is the highest-risk application. Tanno et al. (2025) [8] showed VLM-assisted reports reduced omission errors but introduced hallucinated findings.

Cognitive Priming Risk

A radiologist receiving a VLM-generated draft report receives a primed cognitive frame that alters subsequent pattern recognition, potentially impairing diagnostic accuracy even among experienced physicians [9].

⚠ Devil's Advocate

Contextualise hallucinations against human baselines (3-4% daily error, ~30% retrospective error [15]). If VLMs reduce omission by 40% while hallucinating 2%, net benefit is positive.

Automation Bias

Humans exhibit "comission bias" — preferring action consistent with automated recommendations, especially under cognitive load.

Dratsch et al. (2023) [6]: Incorrect AI suggestions impaired all readers. Inexperienced readers were highly susceptible.

Radiologia Medica (2025) [7]: False-positive AI led to elevated aneurysm suspicion. Reading times were shorter, indicating constrained independent reasoning.

Conceptual Impact: AI False Positives

Illustrative representation synthesizing trends from Dratsch et al. and related literature. Not direct experimental values.

⚠

Devil's Advocate

Laboratory reader studies lack real-world calibration. Real radiologists develop calibrated skepticism over time. The fair comparison is AI with training vs. without.

Governance & Teaching

Technical literacy alone is insufficient. Operational governability is the missing half of the curriculum.

The Teaching Framework

Category	Knowledge Source	Explainability
1. Symbolic AI	Human-encoded rules(Expert systems)	Complete
2. Classical Statistical	Data-learned shallow patterns(Random Forest)	High
3. Deep Statistical	Data-learned hierarchical(CNN, Foundation models)	Low–Moderate

Deployment Checklist

✔

Knowledge Check Quiz

Test your understanding of the core concepts covered in this white paper.

Quiz Complete!

Your score:

Agentic Radiology: The Governance Frontier

Systems that orchestrate multiple steps autonomously pose the most severe governance gap. Existing regulatory frameworks (EU AI Act) were designed for single-task support, not multi-step autonomous agents. The accountability chain is radically unclear (NIST).

Devil's Advocate: Waiting for complete regulatory clarity before engaging with agentic AI deployment is not safety; it is an abdication of professional responsibility to proactively shape governance.

References & Evidence Base

Tavakoli N, Shakeri Z, Gowda V, et al. Generative AI and Foundation Models in Radiology: Applications, Opportunities, and Potential Challenges. Radiology. 2025;317(2):e242961.

Paschali M, Chen Z, Blankemeier L, et al. Foundation Models in Radiology: What, How, Why, and Why Not. Radiology. 2025;314(2):e240597.

Wu C, Zhang X, Zhang Y, Wang Y, Xie W. Towards Generalist Foundation Model for Radiology by Leveraging Web-scale 2D&3D Medical Data. Nature Communications. 2025;16(1).

AI4HI Network. Foundation models for radiology — the position of the AI for Health Imaging (AI4HI) network. 2025.

Springer/JIIM. Large-Scale Foundation Models for Radiological Image Analysis: Clinical Applications, Technical Challenges, and Future Directions. Journal of Imaging Informatics in Medicine. 2026.

Dratsch T, Chen X, Bürkle AC, et al. Automation Bias in Mammography: The Impact of Artificial Intelligence BI-RADS Suggestions on Reader Performance. Radiology. 2023;307:e222176.

Automation Bias in AI-assisted detection of cerebral aneurysms on time-of-flight MR angiography. La Radiologia Medica. 2025.

Tanno R, Barrett DGT, Sellergren A, et al. Collaboration between clinicians and vision-language models in radiology report generation. Nature Medicine. 2025;31(2):599–608.

Hager P, Jungmann F, Holland R, et al. Evaluation and mitigation of the limitations of large language models in clinical decision-making. Nature Medicine. 2024;30:2613–2622.

Zech JR, Badgeley MA, Liu M, et al. Variable generalization performance of a deep learning model to detect pneumonia from chest radiographs: A cross-sectional study. PLOS Medicine. 2018;15(11):e1002686.

Brady AP. Error and discrepancy in radiology: inevitable or avoidable? Insights into Imaging. 2016;8(1):171–182.

European Commission. Regulation (EU) 2024/1689 of the European Parliament and of the Council — Artificial Intelligence Act. Official Journal of the European Union. 2024.

NIST. Artificial Intelligence Risk Management Framework (AI RMF 1.0). National Institute of Standards and Technology. 2023. NIST AI 100-1.

About IRHAI and BeResponsibleAI

The Institute for Responsible Healthcare AI (IRHAI) was established to advance the responsible integration of artificial intelligence into clinical medicine, with a focus on governance frameworks, operational accountability, and clinical education.

© 2026 Institute for Responsible Healthcare AI (IRHAI). Prepared for BeResponsibleAI.com