Recent developments in Artificial Intelligence have increasingly focused on explainability research. The potential of Explainable Artificial Intelligence (XAI) in producing trustworthy computer-aided diagnosis systems and its usage for knowledge discovery are gaining interest in the medical imaging (MI) community to support the diagnostic process and the discovery of image biomarkers. Most of the existing XAI applications in MI are focused on interpreting the predictions made using deep neural networks, typically including attribution techniques with saliency map approaches and other feature visualization methods. However, these are often criticized for providing incorrect and incomplete representations of the black-box models’ behaviour. This highlights the importance of proposing models intentionally designed to be self-explanatory. In particular, part-prototype (PP) models are interpretable-by-design computer vision (CV) models that base their decision process on learning and identifying representative prototypical parts from input images, and they are gaining increasing interest and results in MI applications. However, the medical field has unique characteristics that could benefit from more advanced implementations of these types of architectures. This narrative review summarizes existing PP networks, their application in MI analysis, and current challenges.

Left Ventricle (LV) detection from Cardiac Magnetic Resonance (CMR) imaging is a fundamental step, preliminary to myocardium segmentation and characterization. This paper focuses on the application of a Visual Transformer (ViT), a novel neural network architecture, to automatically detect LV from CMR relaxometry sequences. We implemented an object detector based on the ViT model to identify LV from CMR multi-echo T2* sequences. We evaluated performances differentiated by slice location according to the American Heart Association model using 5-fold cross-validation and on an independent dataset of CMR T2*, T2, and T1 acquisitions. To the best of our knowledge, this is the first attempt to localize LV from relaxometry sequences and the first application of ViT for LV detection. We collected an Intersection over Union ($IoU$) index of 0.68 and a Correct Identification Rate ($CIR$) of blood pool centroid of 0.99, comparable with other state-of-the-art methods. $IoU$ and $CIR$ values were significantly lower in apical slices. No significant differences in performances were assessed on independent T2* dataset ($IoU$ = 0.68, p = 0.405; $CIR$ = 0.94, p = 0.066). Performances were significantly worse on the T2 and T1 independent datasets (T2: $IoU$ = 0.62, $CIR$ = 0.95; T1: $IoU$ = 0.67, $CIR$ = 0.98), but still encouraging considering the different types of acquisition. This study confirms the feasibility of the application of ViT architectures in LV detection and defines a benchmark for relaxometry imaging.

Our work aims to exploit deep learning (DL) models to automatically segment diagnostic regions involved in Alzheimer’s disease (AD) in 18F-fluorodeoxyglucose positron emission tomography (18F-FDG PET) volumetric scans in order to provide a more objective diagnosis of this disease and to reduce the variability induced by manual segmentation. The dataset used in this study consists of 102 volumes (40 controls, 39 with established Alzheimer’s disease (AD), and 23 with established mild cognitive impairment (MCI)). The ground truth was generated by an expert user who identified six regions in original scans, including temporal lobes, parietal lobes, and frontal lobes. The implemented architectures are the U-Net3D and V-Net networks, which were appropriately adapted to our data to optimize performance. All trained segmentation networks were tested on 22 subjects using the Dice similarity coefficient (DSC) and other similarity indices, namely the overlapping area coefficient (AOC) and the extra area coefficient (EAC), to evaluate automatic segmentation. The results of each labeled brain region demonstrate an improvement of 50%, with DSC from about 0.50 for V-Net-based networks to about 0.77 for U-Net3D-based networks. The best performance was achieved by using U-Net3D, with DSC on average equal to 0.76 for frontal lobes, 0.75 for parietal lobes, and 0.76 for temporal lobes. U-Net3D is very promising and is able to segment each region and each class of subjects without being influenced by the presence of hypometabolic regions.

Survival analysis is a statistical approach widely employed to model the time of an event, such as a patient’s death. Classical approaches include the Kaplan–Meier estimator and Cox proportional hazards regression, which assume a linear relationship between the model’s covariates. However, the linearity assumption might pose challenges with high-dimensional data, thus stimulating interest in performing survival analysis using neural network models. In the present work, we implemented a deep Cox neural network (Cox-net) to predict the time of a cardiac event using patient data collected from the Myocardial Iron Overload in Thalassemia (MIOT) project. Cox-net achieved a concordance index (c-index) of 0.812 ± 0.036, outperforming the classical Cox regression (0.790 ± 0.040), and it demonstrated resilience to varying levels of censored patients. A permutation feature importance analysis identified fibrosis and sex as the most significant predictors, aligning with clinical knowledge. Cox-net was able to represent the nonlinear relationships between covariates and maintain reliable survival curve predictions in datasets with a large number of censored patients, making it a promising tool for determining the appropriate clinical pathway for thalassemic patients.