Despite the constant development of new bias mitigation methods for machine learning, no method consistently succeeds, and a fundamental question remains unanswered: when and why do bias mitigation techniques fail? In this paper, we hypothesise that a key factor may be the often-overlooked but crucial step shared by many bias mitigation methods: the definition of subgroups. To investigate this, we conduct a comprehensive evaluation of state-of-the-art bias mitigation methods across multiple vision and language classification tasks, systematically varying subgroup definitions, including coarse, fine-grained, intersectional, and noisy subgroups. Our findings reveal that subgroup choice significantly impacts performance, with certain groupings paradoxically leading to worse outcomes than no mitigation at all. They suggest that observing a disparity between a set of subgroups is not a sufficient reason to use those subgroups for mitigation. Through theoretical analysis, we explain these phenomena and uncover a counter-intuitive insight that, in some cases, improving fairness with respect to a particular set of subgroups is best achieved by using a different set of subgroups for mitigation. Our work highlights the importance of careful subgroup definition in bias mitigation and presents it as an alternative lever for improving the robustness and fairness of machine learning models.

In recent years, the deployment of supervised machine learning techniques for segmentation tasks has significantly increased. Nonetheless, the annotation process for extensive datasets remains costly, labor-intensive, and error-prone. While acquiring sufficiently large datasets to train deep learning models is feasible, these datasets often experience a distribution shift relative to the actual test data. This problem is particularly critical in the domain of medical imaging, where it adversely affects the efficacy of automatic segmentation models. In this work, we introduce DiffuSeg, a novel conditional diffusion model developed for medical image data, that exploits any labels to synthesize new images in the target domain. This allows a number of new research directions, including the segmentation task that motivates this work. Our method only requires label maps from any existing datasets and unlabelled images from the target domain for image diffusion. To learn the target domain knowledge, a feature factorization variational autoencoder is proposed to provide conditional information for the diffusion model. Consequently, the segmentation network can be trained with the given labels and the synthetic images, thus avoiding human annotations. Initially, we apply our method to the MNIST dataset and subsequently adapt it for use with medical image segmentation datasets, such as retinal fundus images for vessel segmentation and MRI images for heart segmentation. Our approach exhibits significant improvements over relevant baselines in both image generation and segmentation accuracy, especially in scenarios where annotations for the target dataset are unavailable during training. An open-source implementation of our approach can be released after reviewing..

While multi-task learning (MTL) has been widely developed for natural image analysis, its potential for enhancing performance in medical imaging remains relatively unexplored. Most methods formulate MTL as a multi-objective problem, inherently forcing all tasks to compete with each other during optimization. In this work, we propose a novel approach by formulating MTL as a multi-level optimization problem, in which the features learned from one task are optimized by benefiting from the other tasks. Specifically, we advocate for a cooperative approach where each task considers the features of others, enabling individual performance enhancement without detriment to others. To achieve this objective, we introduce a novel optimization strategy aimed at seeking flat minima for each sub-problem, fostering the learning of robust sub-models resilient to changes in other sub-models. We demonstrate the advantages of our proposed method through comprehensive parameter and comparison studies on the OrganCMNIST dataset. Additionally, we evaluate its efficacy on three eye-related medical image datasets, comparing its performance against other state-of-the-art MTL approaches. The results highlight the superiority of our method over existing approaches, showcasing its potential for training multi-purpose models in medical image analysis.

Maxime Kayser, Bayar Menzat, Cornelius Emde, Bogdan Bercean, Alex Novak, Abdala Espinosa, Bartlomiej W. Papiez, Susanne Gaube, Thomas Lukasiewicz, Oana-Maria Camburu. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

The success of deep learning on image classification and recognition tasks has led to new applications in diverse contexts, including the field of medical imaging. However, two properties of deep neural networks (DNNs) may limit their future use in medical applications. The first is that DNNs require a large amount of labeled training data, and the second is that the deep learning-based models lack interpretability. In this paper, we propose and investigate a data-efficient framework for the task of general medical image segmentation. We address the two aforementioned challenges by introducing domain knowledge in the form of a strong prior into a deep learning framework. This prior is expressed by a customized dynamical system. We performed experiments on two different datasets, namely JSRT and ISIC2016 (heart and lungs segmentation on chest X-ray images and skin lesion segmentation on dermoscopy images). We have achieved competitive results using the same amount of training data compared to the state-of-the-art methods. More importantly, we demonstrate that our framework is extremely data-efficient, and it can achieve reliable results using extremely limited training data. Furthermore, the proposed method is rotationally invariant and insensitive to initialization.

Deformable image registration, the estimation of the spatial transformation between different images, is an important task in medical imaging. Deep learning techniques have been shown to perform 3D image registration efficiently. However, current registration strategies often only focus on the deformation smoothness, which leads to the ignorance of complicated motion patterns (e.g., separate or sliding motions), especially for the intersection of organs. Thus, the performance when dealing with the discontinuous motions of multiple nearby objects is limited, causing undesired predictive outcomes in clinical usage, such as misidentification and mislocalization of lesions or other abnormalities. Consequently, we proposed a novel registration method to address this issue: a new Motion Separable backbone is exploited to capture the separate motion, with a theoretical analysis of the upper bound of the motions' discontinuity provided. In addition, a novel Residual Aligner module was used to disentangle and refine the predicted motions across the multiple neighboring objects/organs. We evaluate our method, Residual Aligner-based Network (RAN), on abdominal Computed Tomography (CT) scans and it has shown to achieve one of the most accurate unsupervised inter-subject registration for the 9 organs, with the highest-ranked registration of the veins (Dice Similarity Coefficient (%)/Average surface distance (mm): 62%/4.9mm for the vena cava and 34%/7.9mm for the portal and splenic vein), with a smaller model structure and less computation compared to state-of-the-art methods. Furthermore, when applied to lung CT, the RAN achieves comparable results to the best-ranked networks (94%/3.0mm), also with fewer parameters and less computation.

Image corruptions are common in the real world, for example images in the wild may come with unknown blur, bias field, noise, or other kinds of non-linear distributional shifts, thus hampering encoding methods and rendering downstream task unreliable. Image upgradation requires a complicated balance between high-level contextualised information and spatial specific details. Existing approaches to solving the problems are designed to focus on single corruption, which unavoidably results in poor performance when the acquisitions suffer from multiple degradations. In this study, we investigate the possibility of handling multiple degradations and enhancing the quality of images via deblurring, bias field correction, and denoising. To tackle the problems with propagating errors caused by independent learning, we propose a unified and scalable framework, which consists of three special decoders. Two decoders learn artifact attention from provided images thereby generating realistic individual artifact and multiple artifacts on single image; the third decoder is trained towards removing artifact on the synthetic image with multiple corruptions thereby generating high quality image. We additionally provide improvements over previous image degradation synthesis approaches by modelling multiple image degradations directly from data observations. We first create a toy MNIST dataset and investigate the properties of the proposed algorithm. We then use brain MRI datasets to demonstrate our method’s robustness, including both simulated (where necessary) and real-world artifacts. In addition, our method can be used for single/or multiple degradation(s) synthesis by implementing the learned degradation operators in a new domain from a given dataset. The code will be released upon acceptance of the paper.

Rectal tumour segmentation in dynamic contrast-enhanced MRI (DCE-MRI) is a challenging task, and an automated and consistent method would be highly desirable to improve the modelling and prediction of patient outcomes from tissue contrast enhancement characteristics - particularly in routine clinical practice. A framework is developed to automate DCE-MRI tumour segmentation, by introducing: perfusion-supervoxels to over-segment and classify DCE-MRI volumes using the dynamic contrast enhancement characteristics; and the pieces-of-parts graphical model, which adds global (anatomic) constraints that further refine the supervoxel components that comprise the tumour. The framework was evaluated on 23 DCE-MRI scans of patients with rectal adenocarcinomas, and achieved a voxelwise area-under the receiver operating characteristic curve (AUC) of 0.97 compared to expert delineations. Creating a binary tumour segmentation, 21 of the 23 cases were segmented correctly with a median Dice similarity coefficient (DSC) of 0.63, which is close to the inter-rater variability of this challenging task. A second study is also included to demonstrate the method's generalisability and achieved a DSC of 0.71. The framework achieves promising results for the underexplored area of rectal tumour segmentation in DCE-MRI, and the methods have potential to be applied to other DCE-MRI and supervoxel segmentation problems.