Stress-testing medical vision models with images edited by controllable inversion algorithms

Biomedical visual models aim to facilitate the diagnosis, detection, or segmentation of X-ray images. Recently, deep learning has contributed to this task. However, these models are trained on a collection of images that carry biases due to the contexts in which they are usually gathered. During the stress-test process, it is possible to quantitatively measure the performance of these models on instances outside the distribution they were trained on. To this end, it is proposed to edit biomedical images, particularly chest X-rays. Synthetic editing saves the effort of manually collecting medical images or requesting databases that may raise ethical concerns about their origin. The editing is carried out using latent probabilistic diffusion models. The algorithm for inverting diffused images during inference is modified to provide information from the image to be edited and text as condition. In order to avoid entire edition by covering only desired areas, we proposed to contrast conditional and unconditional diffusion outputs as well as iteratively minimizing absolute distance between predicted target and denoised image input. Results show modifications are meet, by providing only extra text specification with no need for masked areas. Finally we measure limitations over ResNet biomedical classification model, allowing to quantify performance with synthetic easy-to-obtain data.