Image variations¶
Data augmentation is a technique used in deep learning to generate additional training data by applying basic transformations such as rotation, intensity operation, noise injection, and filtering. The augmentation addresses the problem of limited or imbalanced datasets, which can lead to overfitting and poor model performance. By increasing the size and diversity of the training set, data augmentation can improve the generalization ability of the model and ultimately result in better predictions.
Data augmentation is powerful in medical imaging, where large datasets are often unavailable. Basic augmentation techniques are easy to apply to medical images, whether 2D or 3D data. By augmenting the data, the model can learn to recognize important features not limited to the original training set, making it more robust and reliable. However, one observed challenge with data augmentation is that it can be difficult to know how many new data points have been generated or the actual value of the augmentation parameters. Therefore, data augmentation is employed and coupled within the sensitivity analysis study in this challenge.
In the case of the challenge, data augmentation is used to create a set of variations of the original dataset to analyze how sensitive the model is to changes in the input data. This can help evaluate the model's reliability and robustness in different scenarios, ultimately leading to more accurate and reliable predictions.
Data augmentation¶
Geometric transformation¶
Geometric transformation is essential in image processing, especially in medical imaging, where the human body's anatomical structures must be accurately visualized. In this context, rotation is a standard geometric transformation that is applied to CT volumes to adjust the orientation of the scan. In this simulation, the rotation is parallel to the z-axis, and the rotation angle is sampled from a normal distribution with a mean of 0 degrees and a standard deviation of 5 degrees. This angle variation is introduced to simulate the rotation of the patient, which can occur due to different pathological conditions that affect the patient's stability on the CT table.
Figure 1: Probability distribution of the angle of rotation variable alpha.
Motion¶
Motion simulation is another critical aspect of medical image processing. The patient's motion during scanning can cause blurring and artifacts in the reconstructed images, leading to diagnostic errors. In this simulation, the patient's movement is simulated by adding a displacement of up to 2 millimeters, a typical range of motion observed in patients. The motion probability distribution is chosen to be uniform, highlighting the equal probability of patient movement.
Figure 2: Probability distribution of the motion variable expressed in millimeters.
Intensity operation¶
Intensity operation is a common technique used to modify the contrast of medical images. In this variation, the intensity operation is applied only to the aortic vessel tree (AVT) to simulate inhomogeneities of the contrast fluid given to the patient. Therefore, the contrast operator can either enhance or diminish the AVT contrast in the thorax volume. The contrast value of the CT volume is raised to the power of gamma, where gamma is calculated from the formulation gamma = exp(beta). The beta value is sampled from a normal distribution with a mean of 0 and a standard deviation 0.05.
Figure 3: Probability distribution of the beta variable from which the contrast gamma is computed.
Noise injection¶
Lastly, noise injection is applied to enhance the robustness of the model to different images' quality. Gaussian noise is implemented with a mean of 0 and a variable standard deviation. The standard deviation of the Gaussian noise is sampled from a Uniform distribution with a minimum value of 0 and a maximum of 0.03. This technique is widely used in medical imaging to account for different noise sources, such as electronic noise, detector noise, and patient motion artifacts.
Figure 4: Probability distribution of the standard deviation of the Gaussian noise.
