Automatic detection of pathological regions in medical images

Medical images are an essential tool in the daily clinical routine for the detection, diagnosis, and monitoring of diseases. Different imaging modalities such as magnetic resonance (MR) or X-ray imaging are used to visualize the manifestations of various diseases, providing physicians with valuable information. However, analyzing every single image by human experts is a tedious and laborious task. Deep learning methods have shown great potential to support this process, but many images are needed to train reliable neural networks. Besides the accuracy of the final method, the interpretability of the results is crucial for a deep learning method to be established. A fundamental problem in the medical field is the availability of sufficiently large datasets due to the variability of different imaging techniques and their configurations.
The aim of this thesis is the development of deep learning methods for the automatic identification of anomalous regions in medical images. Each method is tailored to the amount and type of available data. In the first step, we present a fully supervised segmentation method based on denoising diffusion models. This requires a large dataset with pixel-wise manual annotations of the pathological regions. Due to the implicit ensemble characteristic, our method provides uncertainty maps to allow interpretability of the model’s decisions. Manual pixel-wise annotations face the problems that they are prone to human bias, hard to obtain, and often even unavailable. Weakly supervised methods avoid these issues by only relying on image-level annotations. We present two different approaches based on generative models to generate pixel-wise anomaly maps using only image-level annotations, i.e., a generative adversarial network and a denoising diffusion model. Both perform image-to-image translation between a set of healthy and a set of diseased subjects. Pixel-wise anomaly maps can be obtained by computing the difference between the original image of the diseased subject and the synthetic image of its healthy representation. In an extension of the diffusion-based anomaly detection method, we present a flexible framework to solve various image-to-image translation tasks. With this method, we managed to change the size of tumors in MR images, and we were able to add realistic pathologies to images of healthy subjects.
Finally, we focus on a problem frequently occurring when working with MR images: If not enough data from one MR scanner are available, data from other scanners need to be considered. This multi-scanner setting introduces a bias between the datasets of different scanners, limiting the performance of deep learning models. We present a regularization strategy on the model’s latent space to overcome the problems raised by this multi-site setting.