10 January 2025 in Neural Networks
29 May 2024 in IEEE Transactions on Instrumentation and Measurement
Anomaly detection (AD) in medical images aims to recognize test-time abnormal inputs according to normal samples in the training set. Knowledge distillation based on the teacher-student (T-S) model is a simple and effective method to identify anomalies, yet its efficacy is constrained by the similarity between teacher and student network architectures. To address this problem, in this paper, we propose a T-S model with skip connections (Skip-TS) which is trained by direct reverse knowledge distillation (DRKD) for AD in medical images. First, to overcome the low sensitivity to anomalies caused by structural similarity, we design an encoder-decoder architecture where the teacher network (T-Net) is a pre-trained encoder and the student network (S-Net) is a randomly initialized decoder. During training, the S-Net learns to reconstruct the shallow representations of images from the output of the T-Net, which is called DRKD. Secondly, we introduce skip connections to the T-S model to prevent the S-Net from missing normal information of images at multi-scale. In addition, we design a multi-scale anomaly consistency (MAC) loss to improve the anomaly detection and localization performance. Thorough experiments conducted on twelve public medical datasets and two private medical datasets demonstrate that our approach surpasses the current state-of-the-art by 6.4% and 8.2% in terms of AUROC on public and private datasets, respectively. Code and organized benchmark datasets will be available at https://github.com/Arktis2022/Skip-TS.
16 January 2026 in Biomedical Signal Processing and Control
24 October 2026 in Nature Communications
Autonomous ultrasound robots could improve access to liver disease screening but remain limited by anatomical variability, respiratory motion and bulky hardware. Here, we present a lightweight robotic liver ultrasound system with a 588-g decoupled end-effector controlled by a sonographer-informed AI agent. The robot attaches directly to the abdomen, maintaining stable force and controlled motion during breathing and patient-side disturbances. Trained on sonographer trajectories, the agent integrates spatial, tactile and visual feedback to navigate discontinuous acoustic windows and acquire standard imaging planes. Evaluations in 102 healthy participants and 17 patients yielded clinically interpretable images and anatomical measurements closely matching clinical references. We also deployed the system for hepatic echinococcosis screening in a high-altitude, low-resource setting on the Qinghai-Tibet Plateau and evaluated its performance in dynamic scenarios, including vehicle-based scanning. These findings support portable, patient-mounted robotics for standardized ultrasound beyond specialist settings.
17 June 2026 in Imaging Neuroscience
Fetal brains frequently exhibit anomalies arising from a broad spectrum of etiologies, such as genetic, infectious, hemorrhagic, or hypoxic-ischemic insults, many of which are associated with serious clinical morbidities. Unsupervised anomaly detection, which learns exclusively from normal cases to identify significant deviations from normative patterns without prior knowledge of specific anomaly types, offers a promising approach for automatic diagnosis of such conditions. In particular, recent studies have demonstrated that the absolute age difference (AAD) between predicted gestational age (PGA) of deep learning models from MRI images and biological gestational age (BGA) shows potential for detecting fetal brain anomalies, albeit with limited performance. To enhance anomaly detection capabilities, this study introduces a three-dimensional (3D) Patch-based brain ANomaly Detection framework via Age prediction (PANDA), utilizing the maximum AAD across all patches (MaxAAD) as a biomarker for identifying fetal brain anomalies. Experiments were conducted on MRI data from a large clinical cohort of 1,316 fetuses comprising 711 normal cases and 605 abnormal cases, including 343 with ventriculomegaly (VM), 50 with germinal matrix-intraventricular hemorrhage (GMH-IVH), and 212 with subependymal cysts (SEC). PANDA achieved the best diagnostic performance with an area under the receiver operating characteristic curve (AUROC) of 0.762 and an area under the precision-recall curve (AUPR) of 0.790. Subgroup analysis across the three disease categories further revealed consistently superior performance.
3 November 2026 in Imaging Neuroscience
Deep learning-based super-resolution (SR) techniques offer a powerful tool to enhance the spatial resolution of magnetic resonance imaging (MRI) without increasing scan time. However, most supervised SR methods rely on high-resolution, high-signal-to-noise ratio (SNR) reference data for training, which are difficult to acquire in many clinical settings, such as in time-constrained pediatric imaging. Inspired by the Noise2Noise principle, we propose SRNR, a new training paradigm that enables supervised SR using noisy high-resolution images as references. SRNR was systematically evaluated on both pediatric and adult simulated data across different noise levels of the reference data. SRNR generated super-resolved image volumes with high similarity to the high-resolution references (PSNR: 32.5–35.2 dB) and improved downstream image analysis performance, including tissue segmentation (white matter Dice coefficient: 0.951–0.960) and cortical surface reconstruction (cortical thickness absolute difference: 0.200–0.255 mm). Furthermore, SRNR was shown to be effective on clinically acquired pediatric brain MRI data, achieving improvements consistent with those observed in the simulated experiments. These results demonstrate that SRNR substantially reduces reliance on clean, high-SNR training data and enables effective SR under clinically realistic acquisition constraints, thereby improving the feasibility of deploying deep learning-based SR in pediatric MRI and related applications.