MediScopeDiffusion: A Diffusion Network for 3D

Medical Image Interpretation

Project report submitted to
Indian Institute of Information Technology, Nagpur,
in partial fulfillment of the requirements for the Mini Project - I
at

DEPARTMENT OF COMPUTER SCIENCE AND

ENGINEERING
by
Fahad Ahmad (BT23CSD035)
Ansh Bharadwaj (BT23CSD063)
Anuj Soni (BT23CSD065)
Yash Saini (BT23CSD066)

Under the Supervision of

Mr. Pravin Bhagat

Ms. Nayna Potdukhe

Session Period: July to Dec 2025

भारतीय सूचना प्रौद्योगिकी संस्थान, नागपुर

INDIAN INSTITUTE OF INFORMATION TECHNOLOGY,
NAGPUR
(An Institution of National Importance by Act of Parliament)
Survey No.140,141/1 behind [Link] Wankhede Shetkari Sahkari
Soot
Girni,Nagpur, 441108
 Department of Computer Science and Engineering

Indian Institute of Information Technology, Nagpur

Declaration

We, Fahad Ahmad (BT23CSD035), Ansh Bharadwaj (BT23CSD063),

Anuj Soni (BT23CSD065), Yash Saini (BT23CSD066) , hereby declare that this project work titled
“ MediScopeDiffusion: A Diffusion Network for 3D Medical Image Interpretation ” is carried out by
us in the Department of Computer Science and Engineering of Indian Institute of Information
Technology, Nagpur. The work is original and has not been submitted earlier whole or in part for the
award of any other certification programme at this or any other Institution /University.

Sr. No. Names Signature

1.)
Fahad
Ahmad

2.) Ansh Bharadwaj

3.) Anuj Soni

4.) Yash Saini

Date:
 Certificate

This is to certify that the project titled “ MediScopeDiffusion: A Diffusion Network for 3D Medical Image
Interpretation ”, submitted by Fahad Ahmad (BT23CSD035), Ansh Bharadwaj (BT23CSD063),
Anuj Soni (BT23CSD065), Yash Saini (BT23CSD066) in partial fulfillment of the requirements for the Mini-
Project in CSE department IIIT Nagpur. The work is comprehensive, complete and fit for final evaluation.

Date:

Name of the Supervisor

Designation, Dept, IIIT, Nagpur
 Acknowledgement

We would like to express our sincere gratitude to our project supervisors, Mr. Pravin S. Bhagat and Ms. Nayna
Potdukhe. Their invaluable guidance, constant encouragement, and insightful feedback were instrumental in the
successful completion of this project. Their expertise and mentorship provided a clear direction for our research
and helped us navigate the technical challenges of working with 3D diffusion models.

We are also grateful to the Department of Computer Science and Engineering and the management of the
Indian Institute of Information Technology, Nagpur (IIITN), for providing us with the necessary
infrastructure, computational resources, and a conducive learning environment to carry out this work.

We would also like to acknowledge the collaborative effort of our team members: Fahad Ahmad, Ansh
Bharadwaj, Anuj Soni, and Yash Kumar Saini. This project was a result of our collective hard work,
dedication, and seamless teamwork.

Finally, we owe our deepest gratitude to our parents, family, and friends for their unwavering support, patience,
and encouragement throughout this entire process.
 ABSTRACT

Research on 3D medical image classification is still limited, and most existing methods rely on
very small cropped volumes-such as 32×32×32 or 64×64×64 nodules-rather than full scans.
These approaches work mainly because the region of interest is already isolated and the labels
are clean, but they fail to capture the broader anatomical context needed for robust diagnosis. At
the same time, diffusion models have gained attention for high-quality image generation, yet
their classification potential, especially for 3D medical data, remains largely unexplored.
In this work, we show that diffusion models can be strong 3D feature learners when
provided with structured guidance. We present MediScopeDiffusion, a guided 3D latent-
diffusion framework that uses global scan features and fine-level local features to steer the
denoising process. This guidance enables the model to learn richer spatial representations and
improves its ability to distinguish subtle abnormalities. Experiments on CT data demonstrate
that our approach offers better robustness to noise and improves classification performance over
standard 3D baselines, highlighting a new direction for diffusion-based medical AI.
 Table of Contents
[Link]. Report Content Page No.

1. List of Figures

2. List of Tables

3. List of Symbols and Abbreviations

4. Introduction

5. Literature Review

6.
Background and Motivation

7.
Methodology

8. Results and Discussions

9.
Conclusions

10.
References

LIST OF FIGURES
Figure no. Title Page no.

Figure 3.A.1 Overall architecture

showing image encoding,
saliency-based ROI
extraction, and dense
guidance map generation.

Figure 3.B.1 Diffusion model pipeline

illustrating the forward
training process and
reverse inference process.

Figure 3.C.1 Local–global feature

extraction module
showing slice-level
projection and ROI-based
fusion for generating
and .

LIST OF TABLES
Table No. Title Page No.

Table I (4.A.1) Performance of Diff3Dformer

(Reported Results)

Table II (4.B.1) Performance of X-ray2CTPA

(Reported Results)

5.) List of Symbols, Abbreviations or Nomenclature

 Abbreviation Full Form

AI Artificial Intelligence

CT Computed Tomography

MRI Imaging Magnetic Resonance

CNN Convolutional Neural Network

DDPM Denoising Diffusion Probabilistic Models

ROI Region of Interest

MSE Mean Squared Error

M Dense Guidance Map

F Feature Prior

yl Local Prior

S Saliency Matrix

s Noise term

yg Global Prior

Chapter 1: INTRODUCTION
Research on 3D medical image classification has historically been dominated by nodule-level prediction, where
models operate on small, manually cropped 3D patches-typically 32×32×32 or 64×64×64 voxels-centered around
suspicious lung nodules. This experimental setup is considerably easier because the region of interest (ROI) is
extremely constrained, the nodule location is known a priori, and labels are highly reliable due to radiologist-
generated annotations. Consequently, the model is relieved from interpreting the full anatomical structure of the
lungs and can instead focus on a localized texture classification task.

Popular datasets such as LIDC-IDRI and LUNA16 are specifically designed around this patch-based paradigm.
Many works that claim “state-of-the-art” performance in 3D medical classification—often reporting AUC values
of 0.90 -- 0.95 and accuracies of 90–95%—are evaluated under these simplified conditions. Prominent examples
include Multi-View CNNs [1], 3D Deep Dual-Path Networks [2], Multi-Scale 3D CNNs [3], and Setio et al.’s
CADx system [4]. While these approaches are highly cited, they primarily address the narrow task of pre-
segmented nodule classification rather than the far more challenging problem of full-scan interpretation.

However, restricting models to small ROIs introduces fundamental limitations. Patch-level classifiers lack critical
information such as:

● Global anatomical structure and organ-level context,

● Multi-slice evolution of disease patterns,

● Surrounding tissue characteristics,

● Long-range spatial dependencies across the volume.

As a result, patch-based approaches face several risks:

• Misclassification when abnormalities extend beyond the cropped region,

• Overfitting to local textural cues, reducing real-world generalization,
• Poor robustness under noise, artifacts, and complex pathologies,
• Limited applicability in clinical workflows that require whole-scan reasoning rather than isolated patch
analysis.

These limitations highlight the need to shift from isolated patch classification toward full-volume 3D medical
understanding, where the model learns disease signatures distributed across the entire organ system. Full-scan
classification provides richer contextual information and better aligns with real diagnostic processes, but it also
introduces new challenges: extremely high computational cost, large inter-patient variability, and significantly
higher noise sensitivity.

Parallel to this, diffusion models have recently emerged as powerful generative frameworks capable of learning
rich multi-scale features. Yet, their application to classification-especially for 3D medical volumes-remains
largely unexplored. Existing diffusion research is overwhelmingly focused on image generation, inpainting, or
segmentation, with very limited investigation into their potential as discriminative 3D feature extractors.

In this work, we demonstrate that diffusion models, when augmented with structured priors and anatomically
guided representations, can act as highly effective 3D feature learners. They capture both global volumetric
context and fine-grained local patterns, enabling robust classification across entire scans. This establishes a new
direction for building reliable, context-aware, and clinically deployable 3D medical image analysis systems.

Chapter 2: LITERATURE REVIEW

 Research in 3D medical image analysis has evolved across two major directions: patch-based
nodule classification and whole-volume generative modeling, with limited progress in full-scan
3D classification. This section reviews the key developments in each area and identifies the
limitations that motivate our proposed approach.

A. Patch-Based 3D Lung Nodule Classification

Early literature on computer-aided lung analysis focused heavily on local nodule classification, where models
operate on small cropped regions surrounding suspicious nodules. This paradigm reduces variability and
simplifies the recognition task.

Shen et al. [1] introduced Multi-View CNNs, where a nodule is represented through multiple orthogonal 2.5D
views. Despite strong performance, the model only analyzes restricted local context. Dou et al. [2] proposed a 3D
Deep Dual-Path Network that enhances feature reuse and gradient flow through residual and dense connections.
Similarly, Zhu et al. [3] developed a multi-scale 3D CNN to capture nodule morphology at different resolutions,
while Setio et al. [4] presented one of the earliest end-to-end CADx systems for pulmonary nodules.

Although these methods report very high accuracy and AUC (often exceeding 90%), they rely on pre-segmented
nodule patches (32–64³ voxels). This design ignores global anatomical structure, cannot capture abnormalities
extending beyond the ROI, and lacks applicability for full CT scan interpretation, which is essential in real
clinical workflows. Thus, patch-based methods solve only a simplified sub-task rather than full 3D medical
classification.

B. Foundations of Diffusion Models

Diffusion models have recently emerged as powerful generative frameworks. Sohl-Dickstein et al. [5] introduced
the first mathematically grounded diffusion process based on nonequilibrium thermodynamics. The landmark
work by Ho et al. [6] proposed DDPMs, demonstrating that diffusion can achieve state-of-the-art image
generation by progressively denoising noisy samples.

Subsequent improvements—Nichol and Dhariwal’s cosine schedule [7], Song et al.’s deterministic DDIM
sampling [8], and Rombach et al.’s Latent Diffusion Models (LDMs) [9]—significantly improved sampling speed,
stability, and scalability. These breakthroughs laid the foundation for modern applications such as Stable
Diffusion.

Despite their success, most diffusion research focuses on image synthesis rather than classification, leaving
their discriminative potential underexplored.

C. Diffusion Models in Medical Imaging

Applications of diffusion models in medical imaging are still emerging. Early works such as MedSegDiff [10] and
Med-DDPM [11] applied diffusion for segmentation, showing improved robustness under noise and modality
shifts. Thaler et al. [12] extended latent diffusion to 3D medical volumes, enabling high-quality generative
modeling for CT and MRI. DiffBTS [13] employed diffusion for 3D brain tumor segmentation, while DiffMIC-v2
[14] investigated diffusion-based classification but mainly in 2D imaging settings.

Collectively, these studies reveal that most diffusion-based medical research focuses on reconstruction or
segmentation tasks. Very few extend diffusion models to 3D classification, and none demonstrate their ability to
understand entire volumetric scans using both global context and local anatomical structures.

D. Identified Research Gap

From the reviewed literature, two key limitations emerge:

1. Patch-based nodule classifiers lack global understanding

They only analyze small cropped ROIs, making them unsuitable for disease-level interpretation or
complex abnormalities spread across slices.

2. Diffusion models are underutilized for discriminative 3D learning

Existing works primarily target segmentation or synthesis, with almost no focus on leveraging diffusion-
based representations for full-scan medical classification.

These gaps highlight the need for a framework capable of:

● processing entire 3D volumes,

● capturing both global anatomical context and fine-grained local features, and

● learning robust volumetric representations using diffusion-based denoising.

E. Contribution Motivation
The limitations observed in prior work directly motivate our approach, which integrates structured priors with
diffusion-based feature learning to build a holistic, full-volume 3D medical classification system. By combining
global feature extraction, local saliency-driven guidance, and diffusion modeling, our method aims to address the
shortcomings of both patch-based CNN approaches and existing diffusion architectures.
 Chapter 3: BACKGROUND AND MOTIVATION

A. Diffusion Models: Foundations and Mathematical Formulation

Diffusion models have emerged as one of the most powerful generative modeling techniques in recent years. The
foundational work by Sohl-Dickstein et al. [5] first introduced the idea of gradually adding noise to data and
learning the reverse process. This framework was later popularized and significantly improved by Ho et al. in
their seminal Denoising Diffusion Probabilistic Models (DDPM) paper [6]. Subsequent works such as Improved
DDPM [7], DDIM [8], and Latent Diffusion Models (LDM) [9] further enhanced sampling efficiency and reduced
computational cost.

Forward (Diffusion) Process: The forward process progressively corrupts a clean data sample x₀ into pure noise
x_T using a fixed Markov chain:

(1)

where β_t is a small variance schedule.

By reparameterizing, any noisy sample x_t can be expressed as:

(2)

where

Reverse (Denoising) Process: The denoising model p_θ learns to predict the original noise component:

(3)

The DDPM training objective minimizes the noise-prediction loss:

(4)

This iterative denoising process forces the model to learn deep structural patterns, making diffusion models highly
expressive and robust.

B. Diffusion Models in Medical Imaging: Progress and Limitations

Diffusion models have been adopted in medical imaging primarily for image synthesis, reconstruction, and
segmentation. Examples include MedSegDiff [10] for segmentation, Med-DDPM [11] for medical denoising, as
well as Latent Diffusion for 3D medical scans [12] and DiffBTS for 3D brain tumor segmentation [13].

These studies demonstrate that diffusion models:

 • capture fine structural details in volumetric medical data,
• are highly robust to noise and imaging artifacts,
• outperform GAN-based generative models in stability.

However, very few works attempt to use diffusion models for classification. Most medical diffusion research
focuses on:
• 2D or 3D image generation,
• lesion or organ segmentation,
• synthetic data augmentation,
• modality translation (e.g., MRI → CT).

To the best of our knowledge, the use of diffusion models as discriminative feature learners for 3D medical
classification remains significantly underexplored. Existing classification pipelines rely heavily on CNN-based or
transformer-based backbones that struggle with:
• global anatomical context,
• inter-slice dependencies,
• noise and high similarity between classes,
• computational limitations for full 3D volumes.

This creates a notable research gap: Can diffusion models—with their inherent denoising and structural learning
capabilities—be repurposed for robust 3D classification?

C. Our Contribution: Guided 3D Diffusion for Medical Classification

To address this gap, we propose MediScopeDiffusion, a guided 3D latent diffusion architecture tailored for
medical image classification.

Our contributions are threefold:

• Guidance Vector (M): A spatial map that adaptively determines how much the diffusion model should rely on
global features vs. local ROI features at each voxel location.

• Feature Vector (F): A cross-attention mechanism integrating global anatomical structure with fine-grained
local patterns, providing rich contextual priors during diffusion.

• Guided Latent Diffusion for Classification: Instead of generating images, our diffusion model learns
discriminative 3D representations. By embedding (M, F) inside the denoising process, the model becomes capable
of learning both global and local spatial features. A lightweight classifier head applied on the denoised latent
space yields the final prediction.
Together, these components allow diffusion models to move beyond traditional generative tasks and function as
powerful 3D feature extractors. Experimental results show improved robustness, stronger noise resistance, and
better generalization compared to standard 3D CNN baselines.

Chapter 4: METHODOLOGY

A. High-Level Architecture Overview

 Figure 3.A.1

Figure 3.A.1 illustrates the overall pipeline of the proposed framework. Given a full 3D medical volume

the model extracts two complementary priors:

1. Global prior( ) obtained from the entire scan.

2. Local prior( )derived from the most informative slices and their corresponding Regions of Interest
(ROIs).

These two priors jointly guide the diffusion model.

1) Global Encoder and Global Prior

A lightweight global encoder processes all slices in the scan to capture coarse anatomical context, focusing on:

● Lung shape and structure

● Global intensity distribution

● Large-scale abnormalities

A 1×1 convolutional projection head maps these features into a global prediction vector:
This produces a high-level prior representing the overall anatomical condition.

2) Slice-Level Saliency Maps and ROI Extraction

To identify the most informative slices, we compute slice-wise saliency maps

, where high saliency values indicate richer diagnostic information.

The top-k most informative slices are selected via:

From these selected slices, local patches (ROIs) are cropped around high-saliency regions, enabling the model to
focus on subtle patterns such as nodules, lesions, infiltrations, or focal abnormalities.

3) Local Encoder and Local Prior

The cropped ROIs are passed through a dedicated local encoder designed to capture fine-grained spatial
features that may be absent in global representations.

An attention mechanism aggregates ROI embeddings:

A linear classifier head generates the local prior:

4) Intuition Behind the Two-Prior Design

The two priors play complementary roles:

 ● Global Prior: Captures holistic structures—lung shape, symmetry, and coarse anatomical patterns.

● Local Prior: Captures subtle regions—fine lesions, irregular textures, and small nodules.

● Saliency-driven ROI extraction: Ensures the model focuses only on the most diagnostically critical
slices.

Together, the priors provide strong guidance for the diffusion model, later used to construct the Dense Guidance
Map (M) and Feature Prior (F).

B. Dense Guidance Map (M) and Image Feature Prior (F)

Figure 3.B.1

After obtaining and , two key components are constructed to guide the diffusion process:

1. Dense Guidance Map (M) - voxel-wise spatial guidance

2. Image Feature Prior (F) - semantic feature-level guidance

These priors play a similar role to guided conditioning in diffusion-based classification frameworks such as
DiffMIC-v2.
1) Dense Guidance Map (M)

This map enables every spatial location in the latent space to decide whether to trust the global or local prior.

a) Distance Matrix Construction

We compute a normalized positional distance matrix:

This matrix defines interpolation ratios across spatial positions.

b) Interpolated Prior Map

The vectors and are broadcast into 2D maps and interpolated:

yielding:

Since , the map is symmetric.

c) Intuition

The Dense Guidance Map acts as a soft expert-switching mechanism:

● Voxels with small : rely on global information

● Voxels with large : rely on local cues

2) Image Feature Prior (F)

While MMM provides spatial guidance, F provides semantic guidance.

Let:

● = global image embedding

● = ROI-specific embeddings

These are concatenated and modulated using a learnable gate matrix :

During diffusion, F injects:

● global context

● fine-grained ROI structure

● multi-scale attention-weighted features

directly into the U-Net backbone.

Summary

Both priors-M and F-provide:

● voxel-wise spatial guidance,

● semantic and anatomical priors,

● stable and interpretable diffusion trajectories.

These priors are injected at every U-Net denoising step.

C. Diffusion-Based Training and Classification

 Figure 3.C.1

Once the priors are constructed, they are integrated into the diffusion model during both training and inference.

1) Forward Diffusion (Training Stage)

During training, the clean label vector is gradually perturbed into using the standard DDPM forward
process.

The U-Net receives:

● noisy latent variable

● Dense Guidance Map M

● Image Feature Prior F

and predicts the injected noise :

This encourages the model to learn both semantic and position-aware features.
2) Reverse Diffusion (Inference Stage)

During inference, the model starts from a noisy latent and iteratively removes noise using the learned U-Net.

At each timestep:

● M identifies which regions need global vs. local focus

● F supplies multi-scale contextual features

This produces a clean latent representation .

3) Final Classification

A lightweight classification head is applied to the final denoised latent, producing the predicted class label.

The combination of diffusion-based feature learning, spatial priors, and semantic priors significantly improves
robustness, interpretability, and accuracy, particularly for noisy or complex 3D medical scans.

Chapter 5: RESULTS

Most existing state-of-the-art models for 3D medical image analysis primarily address nodule-level
classification, where a small cropped region around a lung nodule (typically voxels) is labeled as
benign or malignant. Representative examples include Multi-View CNNs [1], 3D Deep Dual-Path Networks [2],
Multi-Scale 3D CNNs [3], and the classical nodule-based approach by Setio et al. [4]. Because these methods
operate on localized, cleanly labeled ROIs, they routinely achieve high performance (90–95% accuracy, AUC
0.90–0.95) on well-curated datasets such as LIDC-IDRI and LUNA16.
However, such results are not directly comparable to our problem setting. These methods solve an easier, patch-
based classification task and do not process the full CT volume. Many “SOTA” claims in the literature are based
on evaluating a single nodule per patient, rather than full-scan, patient-level classification, leading to inflated
performance and increased susceptibility to overfitting.

In contrast, our method performs full-volume 3D classification, integrating global and local contextual signals
using a guided diffusion framework. This makes the task more challenging but substantially more clinically
relevant.

A. Comparison with Diff3Dformer

To contextualize our performance, we compare our results with Diff3Dformer, a recent diffusion–transformer
framework designed for full-scan binary CT classification. Their work is directly comparable to ours, as it also
performs patient-level binary classification on ~600–700 CT scans for pneumonia-related abnormalities.

Table I summarizes the performance reported in their paper. Diff3Dformer achieves its strongest performance
when diffusion guidance is combined with a clustering-based ViT embedding

Table I – Performance of Diff3Dformer (Reported Results)

Method AUC Acc Sens Spec F1

Contrastive + ViT 0.75 0.74 0.68 0.77 0.72

Diffusion + ViT 0.76 0.76 0.68 0.79 0.73

 Contrastive + Clustering ViT 0.78 0.75 0.75 0.74 0.74

Diffusion + Clustering ViT (Diff3Dformer) 0.79 0.77 0.77 0.75 0.76

Table I (4.A.1)

In comparison, our model achieves:

● Accuracy: 75%

● Sensitivity: 95%

● F1-score: 77.55

● AUC: 0.79

Although our accuracy is slightly lower than Diff3Dformer, this difference arises from backbone complexity:
Diff3Dformer relies on a heavy Vision Transformer embedding module, whereas our architecture uses a
lightweight CNN encoder (12–14M parameters).

Despite our significantly smaller backbone, our F1-score (77.55) exceeds Diff3Dformer (76), and our sensitivity
reaches 95%, demonstrating strong abnormality detection capability. This highlights that our guided diffusion
design achieves transformer-level discrimination with a fraction of the computational cost.

Overall, our model achieves a favorable trade-off between performance and efficiency, enabling fast, resource-
friendly full-scan classification.

B. Comparison with X-ray2CTPA

We further compare our method with X-ray2CTPA, a cross-modality diffusion-based pipeline for pulmonary
embolism detection. Their approach synthesizes a 3D CTPA volume from a single 2D chest X-ray using a
diffusion model and classifies the resulting volume using a 3D DenseNet-121 encoder.

Table II – Performance of X-ray2CTPA (Reported Results)

 Metric AUC Acc Sens Spec

X-ray2CTPA (3D DenseNet-121) 0.80 74.0% 73.91% 84.7%

Table I (4.B.1)

Although the cross-modality synthesis is innovative, the reported classification performance remains moderate,
reflecting the difficulty of volumetric diagnosis even with diffusion-assisted reconstruction.

In contrast, our model achieves 75% accuracy and 95% sensitivity, while using a much more compact CNN
encoder (12–14M parameters). Our Dense Guidance Map M and Feature Prior F significantly enhance robustness
without requiring synthetic 3D volume generation or deep, heavy backbones.

This comparison reinforces that true 3D full-volume classification-based on real CT scans rather than
synthesized volumes-remains a challenging but clinically important task, and our method delivers a practical and
effective solution.

C. Comparison with 3D CNN-Based CT Classification

We also compare our approach with the 3D CNN baseline introduced in “Convolutional neural network-based
classification and monitoring models for lung cancer detection: 3D perspective approach.” In their Case 3.1
experiment, a standard 3D CNN was trained on the MosMedData COVID/viral pneumonia dataset (~700 CT
scans) for binary classification.

Their reported performance is:

● Accuracy: 76.44%

● (No sensitivity, specificity, F1-score, or AUC reported.)

This limits the ability to perform a detailed comparison.

In contrast, our model provides a complete metric profile, achieving:

● Accuracy: 75%

● Sensitivity: 95%

● F1-score: 77.55

● AUC: 0.79

Although their accuracy is marginally higher, our significantly better sensitivity and F1-score indicate stronger
abnormality detection. Moreover, our diffusion-guided design captures richer multi-scale structural cues than
standard 3D CNNs, resulting in improved robustness.

Chapter 6: CONCLUSION

In this work, we investigated the effectiveness of diffusion models for 3D medical image classification and
demonstrated that diffusion-guided architectures can operate as strong discriminative learners, not solely as
generative models. The proposed framework integrates global and local anatomical priors using a lightweight
CNN-based embedding module, yet achieves performance comparable to significantly larger transformer-based
and 3D CNN baselines. Unlike many existing state-of-the-art approaches that rely on small, fixed ROI patches,
our method processes the entire 3D CT volume, enabling holistic feature learning analogous to radiological
interpretation.

Despite computational constraints that limited the use of heavier Vision Transformer embeddings or deeper
diffusion backbones, our model achieves competitive accuracy, excellent sensitivity, and a balanced F1-score,
highlighting the strength of guided diffusion mechanisms for volumetric representation learning. These results
underscore that diffusion-driven priors offer a robust pathway for capturing multi-scale structural and semantic
information in full-scan 3D medical data.

For future work, we plan to scale the architecture with stronger encoders, integrate ViT-based feature extractors,
and train on larger and more diverse datasets. These extensions are expected to further enhance classification
performance, generalization capability, and clinical applicability.

Chapter 7: REFERENCES

[1] W. Shen et al., “Multi-view convolutional neural networks for lung nodule classification,” PLoS
ONE, 2017.

[2] Q. Dou et al., “3D deep dual-path network for lung nodule classification,” in MICCAI, 2019.

[3] X. Zhu et al., “Multi-scale 3D convolutional neural network for lung nodule malignancy
classification,” arXiv preprint, 2018.

[4] A. A. A. Setio et al., “Pulmonary nodule detection and classification with 3D convolutional neural
networks,” Medical Image Analysis, 2016.
[5] J. Sohl-Dickstein et al., “Deep unsupervised learning using nonequilibrium thermodynamics,” in
ICML, 2015.

[6] J. Ho et al., “Denoising diffusion probabilistic models,” in NeurIPS, 2020.

[7] A. Nichol and P. Dhariwal, “Improved denoising diffusion probabilistic models,” in ICML, 2021.

[8] J. Song et al., “Denoising diffusion implicit models,” in ICLR, 2021.

[9] R. Rombach et al., “High-resolution image synthesis with latent diffusion models,” in CVPR, 2022.

[10] J. Kim et al., “MedSegDiff: Medical image segmentation with diffusion models,” arXiv preprint,
2022.

[11] J. Wolleb et al., “Med-DDPM: Denoising diffusion models for medical imaging,” arXiv preprint,
2021.

[12] F. Thaler et al., “Latent diffusion models for 3D medical imaging,” arXiv preprint, 2024.

[13] D. Nie et al., “DiffBTS: A lightweight diffusion model for 3D brain tumor segmentation,” Sensors,
2025.

[14] H. Wu et al., “DiffMIC-v2: Medical image classification via improved diffusion network,” IEEE
Transactions on Medical Imaging, 2025.