
Traditional iterative optimization methods are widely used in microwave hyperthermia to concentrate the specific absorption rate (SAR) within the tumor while minimizing undesired heating in healthy tissue. However, variations in treatment conditions require restarting the optimization process, leading to high computational cost and limiting real-time clinical applicability. In this two-dimensional (2D) proof-of-concept work, a gradient-based meta-learning (GML) model is proposed to rapidly optimize the feeding parameters of a fixed antenna system for SAR focusing in simplified microwave hyperthermia study. Model performance is evaluated using the target-to-body SAR ratio (TBR) and the undesired hot-spot-to-target quotient (HTQ). Increasing the number of training tasks improves convergence stability, with the best performance achieved using deterministic sampling (TBR = 11.31 ± 0.71, HTQ = 0.65 ± 0.09). The proposed GML improves TBR by 1.63 and 3.55, and reduces HTQ by 0.02 and 0.24, compared to the phase-only model and phase conjugation method, respectively. Furthermore, compared to Particle Swarm Optimization (PSO), GML achieves comparable focusing performance while reducing computation time from 45.59 s to 1.37 s. Furthermore, the proposed approach was analyzed using 2D central slices of four MRI-derived breast phantoms. Fully converged PSO achieved a higher mean TBR than GML (49.73 ± 14.02 versus 29.93 ± 6.35), whereas GML reduced the optimization time by approximately 35-fold and achieved a higher TBR than PSO under an equal-runtime constraint (RM-PSO) (29.93 ± 6.35 versus 25.89 ± 7.56). In the thermal analysis of representative samples from the four models, GML, fully converged PSO, and RM-PSO achieved comparable mean T90 values above 41 ^∘ C (41.10 ± 0.51 ^∘ C, 41.09 ± 0.47 ^∘ C, and 41.31 ± 0.45 ^∘ C, respectively). GML also yielded a lower mean maximum healthy-tissue temperature (42.56 ± 1.45 ^∘ C) than fully converged PSO (43.29 ± 1.11 ^∘ C) and RM-PSO (43.74 ± 0.38 ^∘ C). These results support GML as a computationally efficient approach for rapid antenna-feeding adaptation within the investigated fixed 2D configuration.
Gastrointestinal (GI) endoscopy is an important imaging modality for the screening and diagnosis of diseases in the GI tract. A lesion tracking technique helps maintain the identity and location of an abnormal region over time, enabling consistent monitoring rather than independent frame-by-frame detection. However, the complex endoscopic environment as non-linear camera displacement, soft tissue deformation, and occlusions are challenges for existing Multi-Object Tracking (MOT) methods. To address these issues, we propose UTR-Track, a robust tracking framework that integrates deep state estimation with adaptive feature learning. The method introduces OSNet-DCN, which embeds Deformable Convolutions into an Omni-Scale Network to extract deformation-robust features, coupled with a dual-tracker architecture, named VTA-TL-UKF. This VTA-TL-UKF utilizes an Unscented Kalman Filter with transfer-learning-based covariance stabilization and a Virtual Trajectory Augmentation module to recover fragmented tracks. Extensive experiments on self-collected clinical sequences and a public Hyper-Kvasir dataset with five types of the lesions such as gastric cancer (GC), esophageal cancer (EC), duodenal bulb ulcer (DBU), reflux esophagitis (RE), and gastritis, demonstrate the effectiveness of the proposed method. On the 18 annotated IGH evaluation sequence, using the primary configuration, pairing the tracker with the hybrid-trained appearance model OSNet-DCN ^* , UTR-Track reaches a Multi-Object Tracking Accuracy (MOTA) of 45.4 https://github.com/shredder0812/UTR-Track .
Accurate analysis of pediatric Fiberoptic Endoscopic Evaluation of Swallowing (FEES) is essential for detecting swallowing abnormalities. Automated segmentation of FEES data remains challenging due to high inter-expert annotation variability, severe class imbalance, and the limited availability of labeled pediatric datasets. This study presents a pediatric FEES segmentation framework and a comparative evaluation of supervised and training-free approaches using sparse annotations from five expert clinicians. We evaluated class-wise U-Net + + models, the recently proposed VM-UNet architecture, and Temporal, a pretrained video object segmentation framework applied in a training-free, in-context learning setting. Six anatomical and pathological classes were considered: laryngeal inlet, vocal cords, penetration, residue, spillage, and salivary pooling. Two complementary experiments were performed: an internal frame-level five-fold cross-validation on 38 videos, and an independent patient-level evaluation in which models trained on the 38 videos were tested on 50 newly annotated videos from different patients. In the internal cross-validation, mean Dice was 0.70, 0.69, and 0.77 for U-Net + + , VM-UNet, and Temporal, respectively. On the independent test set, mean video-level Dice across five classes was 0.45, 0.44, and 0.40 (laryngeal inlet Dice up to 0.67 and presence score up to 0.96). These findings show that models trained on sparse, mostly single-expert annotations can segment anatomical structures in unseen pediatric patients. Supervised models were more accurate for anatomical structures, whereas all approaches performed comparably for swallowing events. Penetration remained the most challenging class, consistent with its lowest inter-expert agreement. Frame-level validation substantially overestimated performance, particularly for retrieval-based in-context models.
Accurate pneumothorax classification requires strong discrimination with low deployment cost. We developed a mask-aware knowledge distillation framework in which a mask-guided ResNet-50 teacher transferred dynamically annealed logits and lesion-weighted multilevel attention to a MobileNetV3-Small student, while masks and teacher-side components were removed at inference. A fixed stratified 20 0.938± 0.009 , and the five-model student ensemble achieved a fixed-test AUC of 0.946 (95 512× 512 . Training-time lesion guidance enabled numerically similar discrimination with substantially lower inference cost. No independent external validation was performed, and generalizability across institutions, acquisition protocols, portable radiography settings, and patient populations remains uncertain; verified patient-level evaluation remains necessary before clinical translation.
Dysarthria is a neurological speech disorder characterized by impaired motor control of speech production, resulting in reduced intelligibility. It commonly arises in the context of neurodegenerative conditions, including amyotrophic lateral sclerosis and cerebral palsy. The development of deep learning systems for dysarthria assessment is hindered by the scarcity of annotated data, which restricts the ability of models to generalize to unseen speakers. This study proposes a cross-corpus diffusion-based data augmentation strategy for improving automatic speech intelligibility classification from dysarthric speech. A Grad-TTS model is trained on dysarthric speakers from the TORGO dataset and used to synthesize the vocabulary of UA-Speech, producing speaker-conditioned samples intended to transfer pathological acoustic traits across corpora. The proposed strategy is evaluated both independently and in combination with conventional signal-level augmentation using ResNet-50 and ResNet-152 classifiers in binary and four-class intelligibility classification tasks. The results indicate that cross-corpus synthetic augmentation improves average performance and that, in most settings, its integration with conventional augmentation yields further gains, suggesting that the two strategies introduce complementary forms of variability.