Region-based deep learning methods to enhance subtle lesion detection

Abstract

This thesis explores deep learning (DL) techniques for analysing medical images, with a particular focus on brain CT and MRI.

We begin by proposing a comprehensive semi-automatic pipeline to tackle the challenges of preparing and standardising a dataset of routinely-collected CT brain scans from the Third International Stroke Trial (IST-3) for DL analysis. Using these scans, we develop a convolutional neural network-based method to detect acute ischemic stroke (AIS) lesions and classify the affected brain side.

To address the challenge of correctly classifying subtle lesions, we introduce the Adversarial Counterfactual Attention (ACAT) framework, which addresses the limitations of traditional CNNs in medical imaging tasks where only small parts of the image are informative. ACAT employs saliency maps to obtain soft spatial attention masks that modulate image features at different scales and increases the baseline classification accuracy of lesions in brain CT scans from 71.39% to 72.55 % and of COVID-19 related findings in lung CT scans from 67.71 % to 70.84%.

We investigate the best way to generate the saliency maps employed in our architecture and propose a way to obtain them from adversarially generated counterfactual images. They are able to isolate the area of interest in brain and lung CT scans without using any manual annotations. In the task of localising the lesion location out of 6 possible regions, they obtain a score of 65.05% on brain CT scans, improving the score of 61.29 % obtained with the best competing method.

ACAT is able to identify where an image should be modified, but not exactly how to modify it to obtain a credible counterfactual. Therefore, we present a weakly supervised method for generating healthy counterfactuals of diseased images and obtaining pixel-wise anomaly maps. This approach combines Denoising Diffusion Probabilistic Models (DDPM) and Denoising Diffusion Implicit Models (DDIM) in a novel way to perform targeted modifications to pathological areas while preserving the rest of the image. The process begins with a saliency map obtained through ACAT, which approximately covers the pathological areas. A diffusion model trained on healthy samples is then employed, using DDPM to modify lesion-affected areas within the saliency map, while DDIM ensures accurate reconstruction of normal anatomy outside these regions.

The two parts are also fused at each timestep, to guarantee the generation of a sample with a coherent appearance and a seamless transition between edited and unedited parts. We compare our approach with alternative weakly supervised methods on the task of brain lesion segmentation, achieving the highest mean Dice and IoU scores among the models considered.