Comparative Analysis of Deep Convolutional Neural Networks for Detecting Medical Image Deepfakes

TL;DR

The findings reveal that ResNet50V2 excels in precision and specificity, whereas DenseNet169 is distinguished by its accuracy, recall, and F1-score, and MobileNetV3Large offers competitive performance, emerging as the swiftest among the considered DCNN models while maintaining a relatively small parameter count.

Abstract

Generative Adversarial Networks (GANs) have exhibited noteworthy advancements across various applications, including medical imaging. While numerous state-of-the-art Deep Convolutional Neural Network (DCNN) architectures are renowned for their proficient feature extraction, this paper investigates their efficacy in the context of medical image deepfake detection. The primary objective is to effectively distinguish real from tampered or manipulated medical images by employing a comprehensive evaluation of 13 state-of-the-art DCNNs. Performance is assessed across diverse evaluation metrics, encompassing considerations of time efficiency and computational resource requirements. Our findings reveal that ResNet50V2 excels in precision and specificity, whereas DenseNet169 is distinguished by its accuracy, recall, and F1-score. We investigate the specific scenarios in which one model would be more favorable than another. Additionally, MobileNetV3Large offers competitive performance, emerging as the swiftest among the considered DCNN models while maintaining a relatively small parameter count. We also assess the latent space separability quality across the examined DCNNs, showing superiority in both the DenseNet and EfficientNet model families and entailing a higher understanding of medical image deepfakes. The experimental analysis in this research contributes valuable insights to the field of deepfake image detection in the medical imaging domain.

Figures (7)

• Figure 1: Sequential scan slices of three patients denoted as (a), (b), and (c) on each row, where the column Original denotes the label provided, and columns $[\pm \textit{i}]$ denote the adjacent slices within a range of i of the Original slice. (a) shows that the tumor starts and ends exactly within the range $[-5, +5]$, (b) shows that the tumor starts and ends within the range $[-3, +4]$, and finally (c) shows that the tumor starts and ends out of the range $[-5, +5]$.
• Figure 2: Full architecture used for our experiments, which consists of a Deep Convolutional Neural Network (DCNN), a global average pooling layer, a Dense (fully-connected) layer of size 512 followed by a size of 256, and finally a Sigmoid activation function indicating a probability of how much an image is real.
• Figure 3: Time recorded per inference step after ten inferences on various Deep Convolutional Neural Networks
• Figure 4: Latent space separability quality between real and fake examples, showing penultimate layer embedding of each Deep Convolutional Neural Network (DCNN). (minor areas of collision circled in green)
• Figure 5: Pearson's correlation matrix showing interrelationships among Deep Convolutional Neural Networks (DCNNs) results.