How to Fool Radiologists with Generative Adversarial Networks? A Visual Turing Test for Lung Cancer Diagnosis

Abstract: Discriminating lung nodules as malignant or benign is still an underlying challenge. To address this challenge, radiologists need computer aided diagnosis (CAD) systems which can assist in learning discriminative imaging features corresponding to malignant and benign nodules. However, learning highly discriminative imaging features is an open problem. In this paper, our aim is to learn the most discriminative features pertaining to lung nodules by using an adversarial learning methodology. Specifically, we propose to use unsupervised learning with Deep Convolutional-Generative Adversarial Networks (DC-GANs) to generate lung nodule samples realistically. We hypothesize that imaging features of lung nodules will be discriminative if it is hard to differentiate them (fake) from real (true) nodules. To test this hypothesis, we present Visual Turing tests to two radiologists in order to evaluate the quality of the generated (fake) nodules. Extensive comparisons are performed in discerning real, generated, benign, and malignant nodules. This experimental set up allows us to validate the overall quality of the generated nodules, which can then be used to (1) improve diagnostic decisions by mining highly discriminative imaging features, (2) train radiologists for educational purposes, and (3) generate realistic samples to train deep networks with big data.

Overview of "How to Fool Radiologists with Generative Adversarial Networks?: A Visual Turing Test for Lung Cancer Diagnosis"

This paper explores the application of Deep Convolutional Generative Adversarial Networks (DC-GANs) in generating realistic samples of lung nodules for the purpose of aiding lung cancer diagnosis. The central premise of this research is to improve the ability of radiologists and diagnostic systems to distinguish between malignant and benign lung nodules by generating high-quality synthetic nodule images. The paper's hypothesis rests on the notion that if generated nodule images are indistinguishable from real ones, they can effectively highlight discriminative features, thereby enhancing diagnostic accuracy.

Methodology

The authors utilize the LIDC-IDRI dataset, which contains annotated CT scans of lung nodules, incorporating a meticulous selection process to ensure a balanced set of benign and malignant nodules. The DC-GAN architecture employed is tailored to the dataset, with a generator that produces 56x56 pixel nodule images and a discriminator that outputs a probability score indicating whether the input image is real or synthesized.

In the study, the network is trained under different conditions: generating solely benign, solely malignant, and a combination of both types of nodules. Various Visual Turing tests were conducted with two radiologists, aiming to assess their ability to distinguish between real and generated nodules, and between malignant and benign samples.

Results

Quantitative outcomes demonstrate that the radiologists were occasionally deceived by the generated images, indicating the high quality of the synthetic nodules. Specifically, the False Recognition Rate (FRR), which measures the percentage of fake nodules perceived as real, was significant, illustrating the proficiency of the generator. Meanwhile, the True Recognition Rate (TRR) varied more extensively between the radiologists, highlighting inter-observer variability.

The paper also provides analysis on the inter-observer agreement for distinction between malignant and benign nodules, as well as for real and generated images. Although the study successfully generated realistic samples that at times fooled skilled practitioners, challenges remain in ensuring that these synthetic images exclusively exhibit features pertinent to either benign or malignant nodules.

Implications and Future Directions

This research holds practical implications for enhancing Computer-Aided Diagnosis (CAD) systems through improved feature extraction from CT images. Furthermore, generated images can serve as a valuable resource for training less-experienced radiologists. Theoretically, the study underscores the potential of GANs in medical image synthesis, inviting further exploration into their capacity to generate 3D volumetric data and their integration into existing diagnostic frameworks.

Future investigations should aim to address the limitations noted, such as improving the ability of generated samples to cleanly exhibit characteristics unique to malignant or benign nodules. Expanding the study with a larger cohort of radiologists and leveraging 3D data representation could further substantiate the technique’s robustness and applicability in clinical settings.

In conclusion, this study demonstrates a notable advancement in medical imaging and diagnostic technology, employing adversarial networks to enhance the interpretative capabilities of radiologists. While promising, the research invites continued refinement and validation to augment its clinical utility and reliability.