GENERATION AND IDENTIFICATION OF DEEPFAKE IN MEDICAL IMAGES: A CASE STUDY ON RADIOLOGY AND ENDOSCOPY

• Main Author: Ayubi, Aula
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/85316
• Institution: Institut Teknologi Bandung

The rapid advancement of telemedicine and digitization in healthcare has significantly improved the efficiency and accessibility of medical services globally. Teleradiology, in particular, has become a crucial tool in clinical radiology practice, enabling seamless transfer of diagnostic images and facilitating remote consultations. Advancements in telemedicine have also significantly impacted endoscopic procedures, allowing for remote diagnosis and collaborative care. However, with the increasing adoption of digital imaging and communication technologies, the potential for malicious manipulation of medical images, known as deepfakes, has emerged as a pressing concern. Fraudulent activities targeting Medicare have resulted in significant financial losses for healthcare facilities. Prior studies and reports have indicated that Medicare incurs losses due to fraudulent or unnecessary claims. Manipulation of radiographs is a serious issue that can be exploited for illegal health insurance claims. Deepfakes, synthetic media created using machine learning algorithms, can convincingly manipulate or generate images, audio, and video. The ability to produce realistic-looking medical images poses significant ethical and practical challenges, including misdiagnosis and fraudulent activities. This study aims to develop a more realistic deepfake detection algorithm for various types of medical images, including gastrointestinal, MRI, CT, and X-ray images. This development is crucial to address the potential dangers of deepfakes in the medical context, such as misdiagnosis and insurance fraud. In this research, BlobGAN is employed as the foundation for creating the training dataset and is modified with the addition of a self-attention layer to generate more realistic deepfake images. Evaluation results demonstrate that the modified Blob Generative Adversarial Network (BlobGAN) generates synthetic images of superior quality compared to other generative models. Validation tests were conducted with medical professionals to assess the realism of the generated images. These synthetic images were also tested to improve the detection accuracy of the YoloV5su model, a compact model frequently utilized in previous research. iv To detect deepfake medical images, several YoloV8 and YoloV5 algorithms were employed, including YoloV5nu, YoloV5su, YoloV8n, YoloV8s, YoloV8m, YoloV8l, and YoloV8x. The use of multiple Yolo versions aimed to evaluate the performance of each algorithm in detecting deepfakes in medical images. Evaluation metrics included recall, precision, and Mean Average Precision (mAP). YoloV5su and Convolutional Block Attention Module (CBAM) demonstrated good performance while maintaining a compact model size. Furthermore, testing the impact of adding deepfake data to polyp detection accuracy showed a contribution of approximately 0.02-0.03 mAP improvement. Overall, this research successfully developed a method for generating realistic synthetic images, particularly radiological and endoscopic images, and assessed their quality using objective metrics such as Inception Score (IS) and Frechet Inception Distance (FID). Additionally, the quality of synthetic images was validated through subjective assessments by medical experts. The YoloV5su-based deepfake detection model developed in this study exhibited good performance in identifying deepfake medical images. The results of this research are expected to contribute to the development of more effective deepfake detection technology and explore the potential utilization of synthetic images to enhance disease diagnosis accuracy.