UE22CS441A –Project Phase – 3

Project Progress Review #3

Project Title : Realtime detection of deepfake
manipulations in medical imaging.
Project ID : 137
Project Guide : Prof. Lenish Pramiee
Project Team : Rahul Lingammanavar -
PES2UG22CS429
Sathwik Gowda KS - PES2UG22CS510
Sowjanya MS - PES2UG22CS572
Pratik Manjunath Naik -
PES2UG22CS905
 Outline

• Abstract and Scope of the Project (1 slide)

• Suggestions from Review – 1 & 2 (1 slide)
• Complete pipeline/Architecture of Project (max 2 slides)
• Implementation Details (for each module) (max 4
slides)
• Project Demonstration (keep the demo video ready)
• Test Plan and Strategy (max 2 slide)
• Results and Discussion (max 2 slides)
• Project Report Draft
• Conclusion and Future Work
• References
 Abstract and Scope

Abstract
• Real-Time Detection System – Develop a CNN-based detector that
identifies manipulated scans
• Secure Clinical Monitoring – Deploy the system over a LAN-based setup
that scans incoming MRIs in real time and alerts clinicians of potential
deepfakes.
• Evaluation & Impact – Assess accuracy, latency, and explainability fidelity,
aiming to improve trust, security, and resilience of AI in healthcare
imaging.
Scope
• Development of a robust pipeline for generating and detecting deepfake
MRI scans.
• Real-time monitoring of medical imaging systems to prevent clinical
misdiagnosis due to fake data.
• Contribution to digital forensics in healthcare by providing datasets and
 Suggestions from Review – 1 & 2

Suggestion 1
• Integrate the deepfake-detection pipeline with existing clinical imaging
infrastructure by connecting the system to the hospital PACS (Picture
Archiving and Communication System).
Suggestion 2
• The system analyzes the top four categories of feature manipulation
commonly seen in deepfake medical images
1) texture and noise inconsistencies
2) anatomical/structural anomalies
3) intensity and histogram shifts
4) high-frequency / edge artifacts
Suggestion 3
• Make the UI clean and clinical by showing only key results upfront—use a
simple layout with minimal colors, clear labels, and collapsible advanced
details.
 Suggestions from Review – 1 & 2

Major Suggestions
1) Texture and Noise Inconsistencies
• Fake images often have smooth or strange-looking areas.
• The natural MRI “grain” or noise pattern is missing or looks different.
• These unusual textures are a major sign of manipulation.
2) Anatomical / Structural Anomalies
• Deepfakes sometimes break normal brain structure.
• Parts of the brain may look misplaced, oddly shaped, or suddenly change
between slices.
• If something looks “not medically possible,” it is usually fake.
3) Intensity and Histogram Shifts
• Real MRIs follow a natural brightness and contrast pattern.
• Fake images may be too bright, too dark, or have sudden brightness
jumps.
• These unusual intensity patterns reveal manipulation
 Complete pipeline/Architecture of Project

Image Input
• MRI scan sent from user/hospital/PACS.
• Pre-Processing
• Resize, normalize, remove noise.

Synthetic MRI Generation

• StyleGAN3 / Diffusion / Progressive GAN used to:
• Remove tumors (if present)
• Add tumors (if absent)

Deepfake Detection Module

• CNN-based classifier checks if MRI is real or manipulated.
• Digital forensics block identifies which GAN model created it.
 Complete pipeline/Architecture of Project

Real-Time Alert System

• Sends notification to doctors when manipulation is detected.
Storage / Logs
• Stores reports, detection results, and metadata
 Data Collection and
Preproccesing
Data Collection
• Collect MRI scans from publicly available medical datasets.
• Include images with tumors and without tumors for balance.
• Organize the dataset into folders based on classes
• Remove corrupted or unusable MRI slices.
• Split data into training, validation, and testing sets.
Pre-Processing
• Convert all MRI slices to the same image format (PNG/JPEG or DICOM →
PNG).
• Resize images to a fixed resolution required by the models (e.g., 256×256
or 512×512).
• Normalize intensity values so all images have similar brightness and
contrast.
• Remove noise and artifacts using smoothing or denoising filters.
 Fake Image Generation
Module
Input MRI Selection
• Take real MRI slices from the dataset (tumor / non-tumor).

Pre-Processing
• Normalize intensity.
• Resize to GAN-compatible resolution (e.g., 256×256 or
512×512).
• Align/crop brain region.

StyleGAN3 / Diffusion / Progressive GAN Training

• Train each model on real MRI dataset.
• Learn tumor and non-tumor distributions.
 Fake Image Generation
Module
Tumor Manipulation
Remove unwanted tumor regions.
Add artificial tumor regions.
Change the size of the tumor.
Change the shape or appearance of the tumor.
Adjust the tumor’s position or intensity to make it look real.
 Detection Model
Module
Detection Model (EfficientNet)
• Input MRI image is given to the EfficientNet model.
• EfficientNet extracts important patterns from the MRI (texture, edges,
abnormalities).
• The model compares these patterns with real MRI characteristics
learned during training.
• It detects signs of manipulation such as unnatural textures,
inconsistencies, or artificial tumor regions.
• EfficientNet classifies the image as real or fake.
• The model runs efficiently and gives fast results suitable for real-time
use.
• The final output is sent to the explainability module (Grad-CAM / SHAP)
for highlighting suspicious areas.
 Real Time Interface Module

Core Functionality
• Receive incoming MRI images/streams (PACS, uploader, API).
• Run detection model (EfficientNet) on each image and return a quick
pass/fail score.
• Show confidence score and model source.
• Alert system is integrated

User Interface / UX
• Simple dashboard with incoming queue, recent alerts, and thumbnails.
• Compact alert card showing patient ID (or anon ID), slice number, score,
and quick-preview.
 Real Time Interface Module

Alerts & Notifications

• Real-time desktop/browser notifications for high-confidence
fakes.
Demo
Video
 Real Time Interface Module

Code

[Link]
 Results and Discussion

• High-Impact Accuracy: Achieved 95.4% overall accuracy on the test set,

confirming the model's strong predictive power.
• Prioritized Patient Safety: Reached a 97.8% Recall (Sensitivity). This is our
most critical metric, proving the system is highly effective at catching fakes
and minimizing dangerous false negatives.
• Ensured Clinical Trust: Secured a 92.3% Precision. This high rate is essential
for usability, as it avoids "alarm fatigue" from false positives and ensures
clinicians trust the alerts.
• Robust & Stable Model: Attained a 96.5% AUC Score, demonstrating the
model is an excellent and reliable classifier across all possible thresholds.
 Timeline – Update on Tasks
• End-to-End System Built: A complete, functional prototype has
been successfully developed, from the user interface to the AI
backend.
• Core Model Complete: The EfficientNet-b0 detection model is
fully trained, tested, and has achieved our target metrics (95.4%
Accuracy, 97.8% Recall).
• Application Complete: The full-stack web application is
operational, upload an image, and receive a real-time analysis.
• Integration Complete: The model is successfully integrated into
the web application, proving the viability of the end-to-end
workflow.
• Next Phase: The project is now transitioning from the
implementation phase to the final robustness enhancement and
publication phase for the 8th semester, as planned.
 Conclusion and Future work

• Addresses Critical Threat: Solves the problem of deepfake

manipulations in medical MRIs.
• High-Performance Model: Built an EfficientNet-b0 model with
97.8% Recall (patient safety) and 92.3% Precision (clinical
trust).
• Unique Dataset: Created a robust hybrid dataset (GAN +
physics-based augmentations) to train the model.
• Enhance Robustness: Train against new threats, including
Diffusion Models and common image compression/noise
attacks.
• Expand Modalities: Apply the detection framework to other
critical areas, such as CT scans and X-rays.
 References
[1] S. Solaiyappan and Y. Wen, “Machine learning based medical image deepfake detection: A comparative
study,” 2024.
[2] A. R. Alsabbagh, “Comparative Analysis of Deep Convolutional Neural Networks for Detecting Medical Image
Deepfake,” King Abdullah University of Science and Technology, 8 Jan. 2024.
[3] Department of Network and Computer Security, “Deepfake Detection using Integrate-Backward-Integrate
Logic Optimization Algorithm with CNN,” State University, USA, 13 Jan. 2023.
[4] D. Lakshmi and D. J. Hemanth, “An Overview of Deepfake Methods in Medical Image Processing for Health
Care Applications,” 2024.
[5] L. Stroebel, M. Llewellyn, T. Hartley, T. S. Ip, and M. Ahmed, “A systematic literature review on the
effectiveness of deepfake detection techniques,” 29 Mar. 2023.
[6] “MedNet: Medical deepfakes detection using an improved deep learning approach,” Nov. 2023.
[7] R. A. Ghadage, V. A. Zinj, and P. P. Sarode, “Deepfake Detection in Medical Images,” 2024.
[8] M. Karaköse, H. Yetiş, and M. Çeçen, “A New Approach for Effective Medical Deepfake Detection in Medical
Images,” IEEE Access, Apr. 2024.
[9] Y. Kim, H. Song, and J. Han, “A Deepfake-Based Deep Learning Algorithm for Medical Data Manipulation
Detection,” 2022.
[10] Jun-Yan Zhu, Taesung Park, Phillip Isola, Alexei A. Efros, “Unpaired Image-to-Image Translation using Cycle-
Consistent Adversarial Networks”
[11] Bagher Sistaninejhad , Habib Rasi, Parisa Nayeri“A Review Paper about Deep Learning for Medical Image
Analysis“
Thank
You