|
In this study, we propose an alternative approach to augmenting a dataset utilizing a technique found in video processing called Video Frame Interpolation (VFI). Unlike traditional methods, with VFI we aim to produce images that are neither mere variations of the original images nor entirely synthetic ones, instead providing a middle ground where the images generated are synthetic temporal variations of the original ones. We propose to use pre-trained VFI networks in conjunction with Transfer Learning to develop specialized models capable of interpolating medical images with enough precision so that a medical specialist would deem them clinically plausible. For this study, we worked with a model developed by Niklaus et al., on cardiac ultrasound videos and images alongside a seasoned cardiologist to provide an expert evaluation on the viability of this technique. Our findings indicate that the results produced by our fine-tuned model can indeed be considered realistic, and depending on the use case, the results of the pre-trained model can also be useful. |