In computer-aided diagnosis (CAD) systems, imaging artifacts may cause misdiagnosis which needs to be prevented from the trustworthiness viewpoint. In this tutorial article, we introduce an outlier-robust signal reconstruction technique that is expected to address this critical problem. We discuss a tradeoff between Tukey’s biweight loss and Huber’s loss functions used widely in robust estimation, and we then present the robust technique based on a mathematical formulation using a weakly convex loss function which is free from the tradeoff issue. This article targets those who are interested in signal processing techniques, aiming to serve as a trigger for studies of misdiagnosis prevention based on the robust signal reconstruction technique.

Image reconstruction, which recovers an original image from incomplete and noisy observed data, is inherently an ill-posed inverse problem. To overcome the ill-posedness, a model-based approach has enabled image reconstruction under complex physical constraints while maintaining mathematical rigor. However, the approach has practical limitations in the representation ability of hand-crafted regularization and parameter tuning. To address these challenges, this article provides an overview of deep unrolling, which constructs neural networks by unrolling an optimization algorithm. In particular, we review the fundamental concepts and recent trends in unrolling networks based on primal-dual splitting (PDS) type algorithms.

Inspired by the success of diffusion models in natural image processing, researches on medical image processing using diffusion models have become increasingly active in recent years. In particular, using pre-trained diffusion models as priors for medical image reconstruction is currently one of the most actively studied topics in this field. When incorporating diffusion models into medical image reconstruction, it is crucial to balance the fidelity to the observed data (e.g., projection data) and the diffusion prior. In this study, we propose a medical image reconstruction method based on an ensemble of diffusion models that automatically optimizes this balance, and we apply the proposed method to sparse-view CT image reconstruction.

Positron emission tomography (PET) image reconstruction successfully reduces noise-induced image degradation by modeling the physical degradation of the observed data. However, no image model has the same level of adaptability as Total Variation in computed tomography (CT), and maximum a posteriori estimation is not widely used. Therefore, researchers are working to build PET image models using deep learning and incorporate them into PET image reconstruction. In this paper, we explain two broadly applicable methods from these studies. The first method incorporates an unsupervised, pretraining-free noise reduction method called Deep Image Prior (DIP) into list-mode PET reconstruction for state-of-the-art PET systems. The second method incorporates a method called Deep Diffuse Image Prior (DDIP), which improves the versatility of diffusion models using DIP. Through these studies, we highlight the importance and appeal of integrating deep learning with PET image reconstruction.

If rehabilitation outcomes in patients with early knee osteoarthritis could be predicted before treatment, it would provide valuable information for selecting appropriate candidates and therapeutic strategies. In this study, we predicted rehabilitation outcomes of early knee osteoarthritis from single x-ray image using multiple convolutional neural network (CNN) models and compared classification accuracy and highlighted features across models by visualizing the basis of prediction with Gradient-weighted Class Activation Mapping (Grad-CAM). From standing anteroposterior x-ray image of 64 patients, 128 cropped images centered on each knee joint were obtained, of which 70％ were used for training and 30％ for validation in a four-fold cross-validation. The training dataset was augmented to comprise 816 images in total. The CNN architecture evaluated were AlexNet, DenseNet-201, Inception-ResNet-v2, NASNet-Large, and ResNet-50. NASNet-Large achieved the highest accuracy of 62.5％, and Grad-CAM highlighted the patella and joint space as key regions associated with the highest classification performance.