Slide 1 – Title

Good morning everyone.

Today I will be presenting a research survey paper titled “Generative AI and Foundation Models in
Medical Imaging.”
This paper focuses on how recent advances in generative artificial intelligence—especially diffusion
models and large foundation models—are transforming the way medical imaging AI systems are
designed, trained, and applied.
Rather than proposing a single new algorithm, this paper presents a broader shift in thinking about
medical AI.

Slide 2 – Agenda
I’ll begin by discussing the background and motivation behind this research.
Then I’ll explain the core problem the paper addresses, followed by its goals and methodology.
After that, I’ll cover the key technical topics such as diffusion models, generative segmentation,
medical language models, and foundation models.
Finally, I’ll discuss the analysis, limitations, future improvements, and conclude with the overall
contribution of the paper.

Slide 3 – Background: Traditional Medical AI

Traditionally, medical image analysis relied on task-specific deep learning models.
For example, one convolutional neural network might be trained to detect tumors, while another
model like U-Net would be trained separately for organ segmentation.
These systems required large amounts of labeled data and worked well only for the specific task they
were trained on.
While they achieved good accuracy, they lacked flexibility and scalability in real clinical
environments.

Slide 4 – Limitations of Traditional Approaches

There are several major limitations in traditional medical AI systems.
First, medical data is limited and highly sensitive.
Second, annotation requires expert radiologists, making labeling slow, expensive, and difficult to
scale.
Additionally, task-specific models often fail to generalize across different hospitals, scanners, or
patient populations.
As a result, building reliable medical AI systems becomes costly and fragmented.

Slide 5 – Rise of Generative AI

In recent years, generative AI has shown remarkable success across multiple domains.
Unlike discriminative models that only predict labels, generative models learn the entire data
distribution and can generate new data samples.
This shift became possible due to large-scale training, self-supervised learning, and advances in model
architectures.
These developments motivated researchers to explore whether similar ideas could be applied to
medical imaging.

Slide 6 – Problem Statement

The core problem addressed by this paper is that current medical imaging AI systems are fragmented,
expensive, and difficult to scale.
Each task requires a separate model and heavily relies on labeled data.
The key question raised by the paper is whether generative AI and foundation models can overcome
data scarcity, reduce annotation dependency, and enable more general and robust medical AI systems.

Slide 7 – Goals of the Paper

The goal of this paper is not to introduce a new model or algorithm.
Instead, it aims to systematically analyze existing generative AI techniques and explain how they can
reshape medical imaging.
The paper also seeks to compare traditional and modern approaches, highlight their strengths and
weaknesses, and propose a roadmap for future research in medical AI.

Slide 8 – Methodology
The authors follow a survey-based methodology.
They review a large body of recent literature on diffusion models, generative segmentation, medical
natural language processing, and foundation models.
The methods are categorized, compared, and analyzed to identify trends, challenges, and future
opportunities.
This approach allows the paper to provide a comprehensive overview rather than experimental results.

Slide 9 – Overview of Topics

The paper covers several key topics.
It starts with the fundamentals of generative versus discriminative AI.
Then it discusses diffusion models and their role in medical image generation.
Next, it explores segmentation as a generative task, followed by medical language models and
multimodal learning.
Finally, it introduces foundation models and their importance in building scalable medical AI systems.

Slide 10 – Discriminative vs Generative AI

Discriminative models focus on learning a direct mapping from input to output, such as from an
image to a label or segmentation mask.
In contrast, generative models learn the full underlying data distribution.
This allows them not only to make predictions but also to generate new images, masks, or text.
This distinction is crucial for handling uncertainty and data scarcity in medical applications.

Slide 11 – Diffusion Models: Core Idea

Diffusion models work through a two-stage process.
In the forward process, noise is gradually added to an image until it becomes pure noise.
In the reverse process, the model learns to remove this noise step by step to reconstruct realistic
images.
This gradual denoising process leads to stable training and high-quality image generation, which is
especially important for medical data.

Slide 12 – Why Diffusion Models Fit Medical Imaging

Medical images require anatomical accuracy and fine structural details.
Diffusion models preserve global structure while refining local details during generation.
Compared to GANs, diffusion models are more stable and easier to control, making them better suited
for sensitive medical imaging tasks.

Slide 13 – Medical Image Generation

In medical imaging, generated images are mainly used for data augmentation and rare disease
simulation.
They help improve the robustness of diagnostic models by exposing them to diverse variations.
However, the paper clearly emphasizes that generated images are not intended for direct clinical
diagnosis and must be used carefully.

Slide 14 – Segmentation Using Diffusion Models

Traditional segmentation models directly predict masks from images.
Diffusion-based segmentation reframes this as a generative process, where segmentation masks are
generated iteratively.
This approach improves boundary accuracy and allows the model to represent uncertainty, which is
important in ambiguous anatomical regions.
Slide 15 – Text Generation in Medicine
Medical AI also involves large amounts of text, such as radiology reports and clinical notes.
Transformer-based language models outperform traditional NLP approaches by capturing long-range
dependencies and contextual meaning.
These models enable tasks like report generation, summarization, and clinical reasoning.

Slide 16 – Medical Language Models

General-purpose language models often lack sufficient medical knowledge.
By training models on medical literature and clinical data, medical language models achieve higher
accuracy and reliability.
The goal is to assist clinicians in documentation and decision-making rather than replacing them.

Slide 17 – Foundation Models

Foundation models are large models trained on massive datasets that can be adapted to many
downstream tasks.
They enable few-shot and zero-shot learning, reducing the need for task-specific training.
This represents a major shift in AI development philosophy.

Slide 18 – Foundation Models in Medical Imaging

In medical imaging, foundation models can handle multiple tasks such as classification, segmentation,
and report generation using a single model.
This significantly reduces annotation cost and improves knowledge transfer across tasks.
Such models form the basis for scalable and general-purpose medical AI systems.

Slide 19 – Analysis and Key Findings

The paper highlights several key findings.
Diffusion models outperform GANs in stability and realism.
Generative segmentation improves robustness and uncertainty handling.
Multimodal models enhance reasoning, and foundation models reduce development cost and improve
scalability.

Slide 20 – What Has Changed

There is a clear paradigm shift from narrow, task-specific supervised models to general-purpose, self-
supervised foundation models.
This change enables more flexible, reusable, and efficient medical AI systems.

Slide 21 – Limitations
Despite their advantages, these approaches face challenges.
Training foundation models requires significant computational resources.
Data privacy, bias, and regulatory approval are also major concerns.
These limitations must be addressed before widespread clinical adoption.

Slide 22 – What Can Be Improved

Future research should focus on improving computational efficiency, developing privacy-preserving
learning methods, and designing better evaluation metrics for generative models.
Human-AI collaboration is also essential.

Slide 23 – How to Improve

The paper suggests national-scale medical datasets, federated learning, multimodal pretraining, and
human-in-the-loop systems as promising solutions.
These approaches aim to balance innovation with safety and ethics.

Slide 24 – Overall Contribution

Overall, this paper provides a unified perspective on generative AI in medical imaging.
It connects advances in vision, language, and generative modeling into a single framework.
More importantly, it acts as a roadmap for future research and system development.

Slide 25 – Conclusion
In conclusion, generative AI and foundation models represent a fundamental shift in medical imaging
AI.
They enable scalable, self-supervised, and generalizable systems that can support clinicians more
effectively while reducing data and annotation constraints.
The paper emphasizes that the future of medical AI lies in building intelligent, generative foundation
systems that collaborate with humans rather than replace them.

Slide 26 – Q&A
Thank you for your attention.
I’ll be happy to answer any questions.