Synthesising computed tomography (CT) images from magnetic resonance images (MRI) plays an important role in the field of medical image analysis, both for quantification and diagnostic purposes. Convolutional neural networks (CNNs) have achieved state-of-the-art results in image-to-image translation for brain applications. However, synthesising whole-body images remains largely uncharted territory, involving many challenges, including large image size and limited field of view, complex spatial context, and anatomical differences between images acquired at different times. We propose the use of an uncertainty-aware multi-channel multi-resolution 3D cascade network specifically aiming for whole-body MR to CT synthesis. The Mean Absolute Error on the synthetic CT generated with the MultiRes network (73.90 HU) is compared to multiple baseline CNNs like 3D U-Net (92.89 HU), HighRes3DNet (89.05 HU) and deep boosted regression (77.58 HU) and shows superior synthesis performance. We ultimately exploit the extrapolation properties of the MultiRes networks on sub-regions of the body.
In pediatric acquired brain injury, heterogeneity of functional response to specific rehabilitation treatments is a key confound to medical decisions and outcome prediction. We aimed to identify patient subgroups sharing comparable trajectories, and to implement a method for the early prediction of the long-term recovery course from clinical condition at first discharge. 600 consecutive patients with acquired brain injury (7.4 years ± 5.2; 367 males; median GCS = 6) entered a standardized rehabilitation program. Functional Independent Measure scores were measured yearly, until year 7. We classified the functional trajectories in clusters, through a latent class model. We performed single-subject prediction of trajectory membership in cases unseen during model fitting. Four trajectory types were identified (post.prob. > 0.95): high-start fast (N = 92), low-start fast (N = 168), slow (N = 130) and non-responders (N = 210). Fast responders were older (chigh = 1.8; clow = 1.1) than non-responders and suffered shorter coma (chigh = −14.7; clow = −4.3). High-start fast-responders had shorter length of stay (c = −1.6), and slow responders had lower incidence of epilepsy (c = −1.4), than non-responders (p < 0.001). Single-subject trajectory could be predicted with high accuracy at first discharge (accuracy = 0.80). In conclusion, we stratified patients based on the evolution of their response to a specific treatment program. Data at first discharge predicted the response over 7 years. This method enables early detection of the slow responders, who show poor post-acute functional gains, but achieve recovery comparable to fast responders by year 7. Further external validation in other rehabilitation programs is warranted.
Background and Objectives: This scoping review explores the current state of the art of AI-based applications in the field of orthopedics, focusing on its implementation in diagnostic imaging and preoperative planning of knee joint procedures. Materials and Methods: The search was carried out using the recognized scholarly databases PubMed, Medline and Embase and was set to identify original research addressing AI applied to imaging in knee diagnosis and surgical planning, written in English and published up to January 2025. Results: The search produced 1612 papers, of which 36 were included in our review. All papers addressed AI applied to common imaging methods in clinical practice. Of these, thirty integrated AI-based tools with X-rays, one applied AI to X-rays to produce CT-like 3D reproductions, and two studies applied AI to MRI. Conclusions: Several AI tools have already been validated for enhancing the accuracy of measurements and detecting additional parameters in a shorter time compared to standard assessments. We expect these may soon be introduced into routine clinical practice to streamline a number of technical tasks and in some cases to replace the need for human intervention.
Zero-shot anomaly detection (ZSAD) offers potential for identifying anomalies in medical imaging without task-specific training. In this paper, we evaluate CLIP-based models, originally developed for industrial tasks, on brain tumor detection using the BraTS-MET dataset. Our analysis examines their ability to detect medical-specific anomalies with no or minimal supervision, addressing the challenges posed by limited data annotation. While these models show promise in transferring general knowledge to medical tasks, their performance falls short of the precision required for clinical use. Our findings highlight the need for further adaptation before CLIP-based models can be reliably applied to medical anomaly detection.