Simplicity bias is a critical challenge in neural net-works since it often leads to favoring simpler solutions and learning unintended decision rules captured by spuri-ous correlations, causing models to be biased and dimin-ishing their generalizability. While existing solutions rely on human supervision, obtaining annotations of the dif-ferent bias attributes is often impractical. To tackle this, we present Debiasify, a novel self-distillation approach that works without any prior information about the nature of biases. Our method leverages a new distillation loss to distill knowledge within a network; from a deep layer where complex, highly-predictive features reside, to a shallow layer where simpler yet attribute-conditioned features are found in an unsupervised manner. In this way, Debiasify learns robust, debiased representations that generalize well across various biases and datasets, enhancing worst-group performance and overall accuracy. Extensive experiments on computer vision and medical imaging benchmarks show the efficacy of our method, significantly outperforming the previous unsupervised debiasing methods (e.g., a 10.13% improvement in worst-group accuracy on Wavy Hair classi-fication in CelebA) while achieving comparable or superior performance to supervised methods. Our code is publicly available at the following link:Debiasify.

Continual Learning (CL) enables neural networks to learn new tasks while retaining previous knowledge. However, most CL methods fail to address bias transfer, where spurious correlations propagate to future tasks or influence past knowledge. This bidirectional bias transfer negatively impacts model performance and fairness, especially in medical imaging, where it can lead to misdiagnoses and unequal treatment. In this work, we show that conventional CL methods amplify these biases, posing risks for diverse patient cohorts. To address this, we propose BiasPruner, a framework that mitigates bias propagation through debiased subnetworks, while preserving sequential learning and avoiding catastrophic forgetting. BiasPruner computes a bias attribution score to identify and prune network units responsible for spurious correlations, creating task-specific subnetworks that learn unbiased representations. As new tasks are learned, the framework integrates non-biased units from previous subnetworks to preserve transferable knowledge and prevent bias transfer. During inference, a task-agnostic gating mechanism selects the optimal subnetwork for robust predictions. We evaluate BiasPruner on medical imaging benchmarks, including skin lesion and chest X-ray classification tasks, where biased data (e.g., spurious skin tone correlations) can exacerbate disparities. Our experiments show that BiasPruner outperforms state-of-the-art CL methods in both accuracy and fairness. Code is available at: BiasPruner.

Deep learning models have achieved remarkable success in medical image classification. These models are typically trained once on the available annotated images and thus lack the ability of continually learning new tasks (i.e., new classes or data distributions) due to the problem of catastrophic forgetting. Recently, there has been more interest in designing continual learning methods to learn different tasks presented sequentially over time while preserving previously acquired knowledge. However, these methods focus mainly on preventing catastrophic forgetting and are tested under a closed-world assumption; i.e., assuming the test data is drawn from the same distribution as the training data. In this work, we advance the state-of-the-art in continual learning by proposing GC2 for medical image classification, which learns a sequence of tasks while simultaneously enhancing its out-of-distribution robustness. To alleviate forgetting, GC2 employs a gradual culpability-based network pruning to identify an optimal subnetwork for each task. To improve generalization, GC2 incorporates adversarial image augmentation and knowledge distillation approaches for learning generalized and robust representations for each subnetwork. Our extensive experiments on multiple benchmarks in a task-agnostic inference demonstrate that GC2 significantly outperforms baselines and other continual learning methods in reducing forgetting and enhancing generalization. Our code is publicly available at the following link: https://github.com/nourhanb/TMI2024-GC2.

Deep learning (DL) models trained to minimize empirical risk on a single domain often fail to generalize when applied to other domains. Model failures due to poor generalizability are quite common in practice and may prove quite perilous in mission-critical applications, e.g., diagnostic imaging where real-world data often exhibits pronounced variability. Such limitations have led to increased interest in domain generalization (DG) approaches that improve the ability of models learned from a single or multiple source domains to generalize to out-of-distribution (OOD) test domains. In this work, we propose BoosterNet, a lean add-on network that can be simply appended to any arbitrary core network to improve its generalization capability without requiring any changes in its architecture or training procedure. Specifically, using a novel measure of feature culpability, BoosterNet is trained episodically on the most and least culpable data features extracted from critical units in the core network based on their contribution towards class-specific prediction errors, which have shown to improve generalization. At inference time, corresponding test image features are extracted from the closest class-specific units, determined by smart gating via a Siamese network, and fed to BoosterNet for improved generalization. We evaluate the performance of BoosterNet within two very different classification problems, digits and skin lesions, and demonstrate a marked improvement in model generalization to OOD test domains compared to SOTA.

Kidney volume is an essential biomarker for a number of kidney disease diagnoses, for example, chronic kidney disease. Existing total kidney volume estimation methods often rely on an intermediate kidney segmentation step. On the other hand, automatic kidney localization in volumetric medical images is a critical step that often precedes subsequent data processing and analysis. Most current approaches perform kidney localization via an intermediate classification or regression step. This paper proposes an integrated deep learning approach for (i) kidney localization in computed tomography scans and (ii) segmentation-free renal volume estimation. Our localization method uses a selection-convolutional neural network that approximates the kidney inferior-superior span along the axial direction. Cross-sectional (2D) slices from the estimated span are subsequently used in a combined sagittal-axial Mask-RCNN that detects the organ bounding boxes on the axial and sagittal slices, the combination of which produces a final 3D organ bounding box. Furthermore, we use a fully convolutional network to estimate the kidney volume that skips the segmentation procedure. We also present a mathematical expression to approximate the 'volume error' metric from the 'Sørensen-Dice coefficient.' We accessed 100 patients' CT scans from the Vancouver General Hospital records and obtained 210 patients' CT scans from the 2019 Kidney Tumor Segmentation Challenge database to validate our method. Our method produces a kidney boundary wall localization error of ~2.4mm and a mean volume estimation error of ~5%.