Air handling units (AHUs) are the primary energy-consuming components within heating, ventilation, and air conditioning (HVAC) systems. Therefore, fault diagnosis of AHUs is crucial for ensuring indoor comfort and improving energy efficiency. Although deep learning-based methods have achieved great success in fault diagnosis of AHUs, existing methods require a large amount of labeled operational data for model training. However, acquiring high-quality labeled data demands substantial domain expertise and prohibitive annotation costs, leaving a vast majority of collected operational data unlabeled and hindering the practical deployment of deep learning-based methods. To overcome this bottleneck, this paper proposes a knowledge and data fusion-driven self-supervised learning (KDFD-SSL) method for the fault diagnosis of AHUs with limited labeled data. Specifically, the top-k important physical fault features are selected according to the prior diagnostic knowledge of AHUs. Meanwhile, a stacked autoencoder (SAE) is designed to extract high-level fault features from raw operational data, capturing complementary fault information to the prior diagnostic knowledge. Furthermore, a Transformer-based encoder is built within a self-supervised framework and pre-trained via knowledge-data fusion, which enables the extraction of discriminative and targeted features from unlabeled data, thereby reducing the reliance on labeled data for downstream diagnosis tasks. Experiments on two AHU datasets demonstrate the effectiveness and superiority of KDFD-SSL over other state-of-the-art fault diagnosis methods, particularly with limited labeled data. The proposed KDFD-SSL method provides a viable pathway for transitioning deep learning-based fault diagnosis from data-rich theoretical scenarios to data-scarce, real-world building energy management applications.

Accurate and flexible indoor temperature forecasts are essential for optimizing heating, ventilation, and air conditioning (HVAC) operation, reducing energy consumption, and participating in demand response programs. Fine-resolution forecasts can be post-aggregated when continuous 5-minute measurements and repeated short-interval inference are available, and this strategy is a strong accuracy baseline. This study instead evaluates a fine-to-coarse direct deployment setting in which a Liquid Neural Network (LNN) is calibrated on 5-minute building data and then evaluated directly at 30-minute and 1-hour supervisory-control intervals without retraining or repeated 5-minute inference for the coarse output. The proposed LNN uses elapsed-time-aware continuous-time state updates to rescale the hidden-state transition across these control intervals. Under the single-step direct coarse-interval inference setting, the LNN achieves up to 0.864 in R2 and 0.017 in coefficient of variation of root mean square error (CV-RMSE), and reduces mean squared error (MSE) by up to 36 % compared with the Transformer-family baselines evaluated under the same direct-inference protocol. In the more demanding 12-step coarse-resolution setting, however, R2 decreases substantially, indicating that direct long-horizon coarse-interval inference remains challenging; low CV-RMSE values in this setting should be interpreted cautiously because indoor temperature has a high mean and narrow variation range. These results should be interpreted separately from the rolling aggregation regime, where post-aggregated 5-minute forecasts provide the strongest point-accuracy baseline. Additional aggregation-aware comparisons show that direct LNN inference provides a lower-complexity deployment option and remains competitive under same-origin horizon-matched 1-hour evaluation. With only 8950 parameters, over three orders of magnitude fewer than typical attention models, the LNN enables fast, on-device inference suitable for edge HVAC controllers. Additional physical validation shows that the full-model response to an indoor-temperature perturbation has a fitted decay time constant of 14.80 min, close to the estimated air-dominant thermal time constant of 12.56 min. Closed-loop surrogate model predictive control (MPC) simulations further show that the LNN improves the comfort-energy tradeoff under coarser control resolutions, including an 11.54% nominal control-energy proxy reduction relative to Transformer at the 30-minute resolution under better below-setpoint comfort.

Nucleus detection and classification (NDC) in histopathology analysis is a fundamental task that underpins a wide range of high-level pathology applications. However, existing methods heavily rely on labor-intensive nucleus-level annotations and struggle to fully exploit large-scale unlabeled data for learning discriminative nucleus representations. In this work, we propose MUSE (MUlti-scale denSE self-distillation), a novel self-supervised learning method tailored for NDC. At its core is NuLo (Nucleus-based Local self-distillation), a coordinate-guided mechanism that enables flexible local self-distillation based on predicted nucleus positions. By removing the need for strict spatial alignment between augmented views, NuLo allows critical cross-scale alignment, thus unlocking the capacity of models for fine-grained nucleus-level representation. To support MUSE, we design a simple yet effective encoder-decoder architecture and a large field-of-view semi-supervised fine-tuning strategy that together maximize the value of unlabeled pathology images. Extensive experiments on three widely used benchmarks demonstrate that MUSE effectively addresses the core challenges of histopathological NDC. The resulting models not only surpass state-of-the-art supervised baselines but also outperform generic pathology foundation models.

The global rise in steatotic liver disease poses a significant public health challenge. While non-contrast computed tomography scans hold promise for opportunistic detection of steatotic liver disease, their potential for staging and risk assessment remains underexplored. Here we present a multimodal AI model trained on a large dataset, comprising of (n=968) histopathologically and (n=1103) radiologically confirmed cases, validated against both histology (n=660) and MRI-PDFF (n=375) gold standards, demonstrating high accuracy in detecting mild to severe steatosis (AUC: 0.904-0.929) and clinically significant fibrosis (AUC: 0.824-0.888). Furthermore, integrating the model into the standard clinical pathway improves primary risk screening in a retrospective patient cohort (n=1192), identifying 36% more patients at risk of fibrosis progression. Using Cox proportional hazard model, we observe that the intermediate-high risk patients identified by the optimized clinical pathway exhibits a significantly higher incidence of cirrhosis (hazard ratio: 5.54: 2.69-11.42), showcasing the model's potential for early detection and management of steatotic liver disease.

Fault diagnosis in heating, ventilation, and air-conditioning (HVAC) systems is essential for intelligent energy management. However, extreme class imbalance often degrades diagnostic performance. Existing data augmentation methods typically assume a fixed and known set of fault categories; when new fault categories emerge, these methods may become ineffective. To address this challenge, we propose an adaptive augmentation framework (ADAF) based on a meta-diffusion model. Specifically, we develop a meta-diffusion denoising implicit model (MDDIM) that is meta-trained on known fault categories and can rapidly adapt to imbalanced datasets containing new fault categories, enabling the generation of realistic fault samples. To better capture the distribution of an emerging fault category from scarce data, we further design a prototype-similarity-based strategy to initialize the class condition vector and integrate it with a conditional noise prediction network (CNPN) to improve generation quality. Extensive experiments on three HVAC datasets demonstrate that the proposed method consistently outperforms seven representative augmentation baselines. Across different imbalance settings, MDDIM improves the F1 score by 2.75%-21.75% after augmentation, providing an effective solution for fault diagnosis under extreme imbalance with emerging fault categories.

Lymph node (LN) assessment is an essential task in the routine radiology workflow, providing valuable insights for cancer staging and treatment planning. Identifying scatteredly-distributed and low-contrast LNs in 3D CT scans is highly challenging, even for experienced clinicians. Previous lesion and LN detection methods demonstrate the effectiveness of 2.5D approaches (i.e., using 2D backbone with multi-slice inputs), leveraging pretrained 2D model weights and showing improved accuracy as compared to separate 2D or 3D detectors. However, slice-based 2.5D detectors do not explicitly model inter-slice consistency for LN as a 3D object, requiring heuristic post-merging steps to generate final 3D LN instances, which can involve tuning a set of parameters for each dataset. In this work, we formulate 3D LN detection as a slice-by-slice tracking task along the z-axis and propose LN-Tracker, a novel LN tracking transformer, for joint end-to-end detection and 3D instance association. Built upon a DETR-based detector, LN-Tracker decouples transformer queries into distinct track and detection groups with independent matching, enabling comprehensive LN detection while maintaining trajectory consistency. A masked attention mechanism further separates learning between these query groups, and a similarity loss promotes robust inter-slice LN association, particularly in low-contrast scenarios. Extensive evaluation on four LN datasets shows LN-Tracker's superior performance, with at least ${2}.{49}\%$ gain in average sensitivity when compared to top 3D/2.5D/tracking detectors. Further validation on public lung nodule and prostate tumor detection tasks confirms the generalizability of LN-Tracker as it achieves top performance on both tasks. Code is available at https://github.com/alibaba-damo-academy/LN-Tracker.

Fault diagnosis of heating, ventilation, and air conditioning (HVAC) systems is pivotal for energy conservation and emissions reduction in smart-city energy infrastructure. Although existing HVAC fault diagnosis methods have advanced considerably, most of them require large amounts of fault data and assume identical label spaces between the training and test sets. With limited fault samples and the presence of unknown faults, conventional frameworks often experience marked performance degradation. To address these challenges, a few-shot unknown fault diagnosis framework with entropy-based uncertainty guidance (FEUG) is proposed for HVAC systems. First, to enhance the representational capacity and reusability of features, a feature association fusion module (FAFM) is developed. Building on FAFM, a multi-scale feature interaction association network (MFIAN) is designed to achieve correlation fusion across features of different scales, thereby strengthening joint attention to global and local information. Second, an adaptive contrastive learning module, an entropy-assisted classification module, and a prototype similarity module are proposed to improve intra-class compactness and inter-class separability. Lastly, by integrating entropy-based uncertainty with the few-shot learning mechanism, the FEUG significantly enhances performance in HVAC few-shot unknown fault diagnosis tasks. Comprehensive experiments and ablation analyses conducted on three HVAC datasets and against seven comparative methods demonstrate that FEUG consistently achieves superior accuracy across various scenarios (SC). Specifically, FEUG attains average F1 scores of 69.92%, 82.39%, and 78.48% on SZVAV (SC2), SZCAV (SC8), and chiller (SC12), respectively, exceeding the second-best models by 3.24, 5.39, and 2.35 percentage points.

Automated whole-body lesion segmentation in 18F-FDG PET/CT images marks a pivotal breakthrough in oncological diagnostics, substantially improving the accuracy and efficiency of tumor burden assessment. Manual segmentation is often plagued by significant interobserver variability, underscoring the necessity for automated solutions. The synergistic combination of PET’s exceptional sensitivity for detecting metabolic activity with CT’s anatomical precision renders accurate segmentation crucial for achieving quantitative and reproducible clinical workflows. However, current methodologies frequently grapple with challenges such as over-segmentation or under-segmentation, inadvertently delineating normal tissues with elevated uptake or neglecting lesions characterized by subtle intensity variations, primarily due to a lack of integrated metabolic and anatomical insights. To address these limitations, we present a novel framework that adeptly integrates clinical expertise regarding anatomical and metabolic cues to refine PET/CT lesion segmentation. Our innovative mixture-of-experts (MoE) based interpretable fusion module skillfully merges complementary modality information while explicitly elucidating the pixel-level contributions of each modality to the final segmentation outcome. Rigorous evaluations across three in-domain benchmarks and two external datasets demonstrate our model’s superior segmentation performance and generalizability. Furthermore, our visualizations provide compelling insights into the pivotal role each modality plays in the decision-making process, highlighting our approach’s transformative potential in enhancing PET/CT lesion segmentation. Building on this foundation, we further validated the prognostic significance of the features extracted from our proposed framework in the context of PET/CT-based prognosis predictions.

Accurate tumor analysis is central to clinical radiology and precision oncology, where early detection, reliable lesion characterization, and pathology-level risk assessment directly guide diagnosis, staging, and treatment planning. Chain-of-Thought (CoT) reasoning is particularly critical in this setting, as it enables stepwise interpretation from imaging findings to clinical impressions and pathology-level conclusions, ensuring traceability and reducing diagnostic errors. Here, we target the clinical tumor analysis task and build a large-scale benchmark that operationalizes a multimodal reasoning pipeline, spanning findings, impressions, and pathology predictions. We curate TumorCoT, a large-scale dataset of 1.5M CoT-labeled VQA instructions paired with 3D CT scans, with step-aligned rationales and cross-modal alignments along the “findings → impression → pathology” trajectory, enabling standardized evaluation of both final accuracy and reasoning consistency. We further propose TumorChain, a multimodal interleaved reasoning framework that tightly couples 3D imaging encoders, clinical text understanding, and organ-level vision-language alignment. Through cross-modal alignment and iterative interleaved causal reasoning, TumorChain grounds visual evidence, aggregates conclusions, and issues pathology predictions after multiple rounds of self-refinement, improving traceability and reducing hallucination risk. TumorChain demonstrates consistent gains over strong unimodal and pipeline baselines in lesion detection, impression quality, and pathology classification, and successfully generalizes to the public DeepTumorVQA benchmark. Ablations validate the key contributions of interleaved reasoning and clinical CoT. Clinically, these advances lay the groundwork for reliable, interpretable tumor assessment to support real-world decision-making. We release the task, benchmark, and evaluation protocol to advance safe, explainable, and reproducible multimodal reasoning for high-stakes tumor analysis. Our project is available at https://anonymous.4open.science/r/TumorChain-D6E6.

Background Accurate preoperative identification of pathologic extranodal extension (ENE) at CT is essential for precise treatment decisions in laryngeal and hypopharyngeal squamous cell cancer (LHSCC). However, human interpretation of ENE is neither reliable nor reproducible. Purpose To develop and evaluate the diagnostic performance of a new deep learning tool, DeepENE, in detecting metastatic and ENE lymph nodes on preoperative CT scans in patients with LHSCC in a multicenter cohort. Materials and Methods In this retrospective study, patients with LHSCC from Zhongshan Hospital, Fudan University (April 2011-August 2022), were included in training, validation, and internal test sets to develop DeepENE. For the reference standard, lymph nodes were segmented on CT scans and labeled for metastasis and ENE status based on pathologic findings. DeepENE was tested using three external cohorts of patients with LHSCC (external test sets 1-3) and one external cohort of patients with oral squamous cell carcinoma. The primary diagnostic metric was the area under the receiver operating characteristic curve (AUC). The performance of DeepENE was compared with that of five board-certified head and neck cancer specialists using the DeLong method. Results Overall, 289 patients with LHSCC with 1954 pathologically confirmed lymph nodes were evaluated. DeepENE achieved an AUC of 0.93 for ENE diagnosis in the internal test set under fivefold cross-validation, and AUCs of 0.96, 0.87, and 0.90 in external test sets 1, 2, and 3, respectively. DeepENE outperformed the five experts, especially in early-stage ENE detection in external test set 2 (AUC of 0.87 for DeepENE vs mean AUC of 0.66 for readers; P < .001). In external test set 1, DeepENE maintained a high sensitivity of 97% at specificity of 90%, compared with experts' mean sensitivity of 77% (P = .003). In external test sets 2 and 3, DeepENE had sensitivity of 78% and 80%, compared with experts' mean sensitivity of 36% (P < .001) and 46% (P < .001), respectively. Conclusion DeepENE accurately detected ENE on preoperative CT scans in patients with LHSCC and outperformed head and neck cancer specialists. © RSNA, 2026 Supplemental material is available for this article.

Liver malignancies are frequently evaluated on contrast-enhanced computed tomography (CE-CT), but missed or delayed diagnoses remain a clinically important challenge in high-volume, real-world radiology workflows, highlighting the need for scalable diagnostic safety net approaches. To address this, we developed the Liver DiagnOsis Network (LiON), a CE-CT-based artificial intelligence (AI) system that supports flexible multiphase processing, clinical data integration and workflow-compatible liver malignancy diagnosis. LiON was trained on 6,443 patients and retrospectively validated across 22,251 patients from multicenter and real-world cohorts. LiON achieved high performance for malignancy diagnosis, with an area under the receiver operating characteristic curve (AUC) of 0.975 (95% confidence interval (CI): 0.971-0.979), and maintained robust performance in real-world cohorts and among patients with hepatic steatosis (AUC 0.971, 95% CI: 0.952-0.985) and cirrhosis (AUC 0.924, 95% CI: 0.901-0.946). We then conducted a single-arm trial in 10,333 patients in routine clinical practice, in which LiON functioned as an additional AI reader within the existing clinical workflow. The trial met its primary endpoint, defined as an AUC for malignancy diagnosis with the lower bound of the 95% CI exceeding 0.900, achieving an AUC of 0.952 (95% CI: 0.942-0.961). Secondary outcomes demonstrated that AI-human collaboration identified 51 previously overlooked lesions (15 malignancies) and triggered 37 amended radiology reports, 22 multidisciplinary team escalations and clinical management changes in a subset of patients. These findings suggest that AI, when deployed as a workflow-compatible diagnostic support, may help reduce missed or delayed diagnoses and guide clinical interventions. Nevertheless, further evidence from prospective comparative studies across diverse healthcare systems is warranted to assess effects on clinical outcomes. ClinicalTrials.gov identifier: NCT07153783 .

Esophageal cancer is one of the most lethal cancers, with 5-year survival rate of only 20%. Patient outcomes can vary significantly even though they are at the same cancer stage and receive similar treatments. Accurate prognostic prediction for esophageal cancer patients is highly desired to receive personalized precise treatment. Nevertheless, there are very few automated methods yet to fully exploit the preoperative contrast-enhanced computed tomography (CE-CT) imaging for assessing esophageal cancer prognosis. In addition to image patterns, important prognostic factors should encompass tumor size and location, as well as lymph nodes (LNs) involvement, including features such as LN number, size, spatial distribution, and their proximity to tumor. Considering these complexities, we propose a novel Tumor and LN Context-Geometry network for the preoperative prediction of esophageal cancer survival in CE-CT images. Specifically, we (1) focus on learning survival patterns of CT texture via co-attention context modeling at most informative regions, i.e., automatically segmented tumor, LNs and LN-stations; and (2) integrate tumor and LN anatomical and spatial associations into neural geometry modeling for a comprehensive learning of metastatic involvement and tumor invasion to adjacent structures. Empirical studies show our presented framework can improve overall survival prediction performances compared with existing state-of-the-art survival analysis methods, and evidently suggest that incorporating these findings into the existing esophageal cancer staging system would add its clinical values.

The integration of Building-Integrated Photovoltaic and Battery (BIPVB) systems is a critical pathway toward carbon-neutral buildings, yet their optimal control is challenged by the stochastic nature of renewable generation and complex load dynamics. While Deep Reinforcement Learning (DRL) has emerged as a promising model-free control approach, traditional value-based and policy-gradient methods often suffer from training instability and lack explicit long-horizon optimization capabilities. This study explores the application of the Decision Transformer (DT) to Building-Integrated Photovoltaic and Battery (BIPVB) control by reformulating the task as a return-conditioned sequence modeling problem. By leveraging the Transformer architecture, the proposed method treats the control problem as a supervised learning task, predicting optimal actions based on historical states and desired future returns (Return-to-Go). We conduct a comprehensive benchmark using the CityLearn dataset, comparing DT against a Rule-Based Controller (RBC) and six state-of-the-art DRL algorithms, including SAC, PPO, and TD3. Simulation-based benchmark results show that the DT agent achieves the lowest operational cost across the tested scenarios, yielding an 18.5% cost reduction compared to RBC and outperforming the strongest DRL baseline. Furthermore, the DT shows competitive and relatively consistent performance under varying initial battery conditions. These results suggest that sequence-modeling approaches are a promising direction for offline building energy control and merit further investigation in broader settings.

The accurate diagnosis and segmentation of tumors in contrast-enhanced Computed Tomography (CT) are fundamentally driven by the distinctive hemodynamic profiles of contrast agents over time. However, in real-world clinical practice, complete temporal dynamics are often hard to capture by strict radiation dose limits and inconsistent acquisition protocols across institutions, leading to a prevalent missing modality problem. Existing deep learning approaches typically treat missing phases as absent independent channels, ignoring the inherent temporal continuity of hemodynamics. In this work, we propose Time Attenuated Representation Disentanglement (TARDis), a novel physics-aware framework that redefines missing modalities as missing sample points on a continuous Time-Attenuation Curve. We first hypothesize that the latent feature can be disentangled into a time-invariant static component (anatomy) and a time-dependent dynamic component (perfusion). We achieve this via a dual-path architecture: a quantization-based path using a learnable embedding dictionary to extract consistent anatomical structures, and a probabilistic path using a Hemodynamic Conditional Variational Autoencoder to model dynamic enhancement conditioned on the estimated scan time. This design allows the network to infer missing hemodynamic features by sampling from the learned latent distribution. Extensive experiments on a large-scale multi-modal private abdominal CT dataset (2,282 patients) and two public datasets demonstrate that TARDis significantly outperforms state-of-the-art incomplete modality frameworks. Notably, our method maintains robust diagnostic performance even in extreme data-sparsity scenarios, highlighting its potential for reducing radiation exposure while maintaining diagnostic precision.

Assessing lymph node LN metastasis in CT is critical for esophageal cancer treatment planning. While clinical criteria are commonly used, the diagnostic accuracy is low with sensitivities ranging from 39.7% to 67.2% in previous studies. Deep learning would have the potential to improve it by learning from large-scale accurately labeled data. However, from the surgical procedure in LN dissection, pathological report only indicates the number of dissected LNs in each lymph node station (LN-station) with the number of metastatic ones found in the respective LN-station. So, it is difficult to establish one-to-one pairing between LN instances observed in CT and their metastasis status confirmed in the pathological report. In contrast, gold reference labels on LN-station metastasis can be readily retrieved from pathology reports at scale. Hence, instead of distinguishing LN instance metastasis, we directly classify LN-station metastasis using pathology-confirmed station labels. We first segment mediastinal LN-stations automatically to serve as input for classification. Then, to improve classification performance, we automatically segment all visible LN instances in CT and design a new LN prior-guided attention loss to explicitly regularize the network to focus on regions of suspicious LNs. Furthermore, considering the varying appearances and contexts of different LN-station, we propose a station-aware mixture-of-experts module, where the expert is trained to specialize in a group of LN-stations by learning to route each LN-station group tokens to the corresponding expert. We conduct fivefold cross-validation on 1,153 esophageal cancer patients with CT and pathology reports (the largest study to date), and our method significantly outperforms state-of-the-art approaches by 2.26% in AUROC.

Identifying anatomical structures (e.g., lesions or landmarks) is crucial for medical image analysis. Exemplar-based landmark detection methods are gaining attention as they allow the detection of arbitrary points during inference without needing annotated landmarks during training. These methods use self-supervised learning to create a discriminative voxel embedding and match corresponding landmarks via nearest-neighbor searches, showing promising results. However, current methods still face challenges in (1) differentiating voxels with similar appearance but different semantic meanings (e.g., two adjacent structures without clear borders); (2) matching voxels with similar semantics but markedly different appearance (e.g., the same vessel before and after contrast injection); and (3) cross-modality matching (e.g., CT-MRI landmark-based registration). To overcome these challenges, we propose a Unified framework for learning Anatomical Embeddings (UAE). UAE is designed to learn appearance, semantic, and cross-modality anatomical embeddings. Specifically, UAE incorporates three key innovations: (1) semantic embedding learning with prototypical contrastive loss; (2) a fixed-point-based matching strategy; and (3) an iterative approach for cross-modality embedding learning. We thoroughly evaluated UAE across intra- and inter-modality tasks, including one-shot landmark detection, lesion tracking on longitudinal CT scans, and CT-MRI affine/rigid registration with varying fields of view. Our results suggest that UAE outperforms state-of-the-art methods, offering a robust and versatile approach for landmark-based medical image analysis tasks. Code and trained models are available at: https://github.com/alibaba-damo-academy/self-supervised-anatomical-embedding-v2.

Histopathology plays a critical role in medical diagnostics, with whole slide images (WSIs) offering valuable insights that directly influence clinical decision-making. However, the large size and complexity of WSIs may pose significant challenges for deep learning models, in both computational efficiency and effective representation learning. In this work, we introduce Pixel-Mamba, a novel deep learning architecture designed to efficiently handle gigapixel WSIs. Pixel-Mamba leverages the Mamba module, a state-space model (SSM) with linear memory complexity, and incorporates local inductive biases through progressively expanding tokens, akin to convolutional neural networks. This enables Pixel-Mamba to hierarchically combine both local and global information while efficiently addressing computational challenges. Remarkably, PixelMamba achieves or even surpasses the quantitative performance of state-of-the-art (SOTA) foundation models that were pretrained on millions of WSIs or WSI-text pairs, in a range of tumor staging and survival analysis tasks, even without requiring any pathology-specific pretraining. Extensive experiments demonstrate the efficacy of PixelMamba as a powerful and efficient framework for end-toend WSI analysis.

Segment anything model (SAM) demonstrates strong generalization ability on natural image segmentation. However, its direct adaptation in medical image segmentation tasks shows significant performance drops. It also requires an excessive number of prompt points to obtain a reasonable accuracy. Although quite a few studies explore adapting SAM into medical image volumes, the efficiency of 2D adaptation methods is unsatisfactory and 3D adaptation methods are only capable of segmenting specific organs/tumors. In this work, we propose a comprehensive and scalable 3D SAM model for whole-body CT segmentation, named CT-SAM3D. Instead of adapting SAM, we propose a 3D promptable segmentation model using a (nearly) fully labeled CT dataset. To train CT-SAM3D effectively, ensuring the model's accurate responses to higher-dimensional spatial prompts is crucial, and 3D patch-wise training is required due to GPU memory constraints. Therefore, we propose two key technical developments: 1) a progressively and spatially aligned prompt encoding method to effectively encode click prompts in local 3D space; and 2) a cross-patch prompt scheme to capture more 3D spatial context, which is beneficial for reducing the editing workloads when interactively prompting on large organs. CT-SAM3D is trained using a curated dataset of 1204 CT scans containing 107 whole-body anatomies and extensively validated using five datasets, achieving significantly better results against all previous SAM-derived models.

Positron emission tomography/computed tomography (PET/CT) is essential in oncology, yet the rapid expansion of scanners has outpaced the availability of trained specialists, making automated PET/CT report generation (PETRG) increasingly important for reducing clinical workload. Compared with structural imaging (e.g., X-ray, CT, and MRI), functional PET poses distinct challenges: metabolic patterns vary with tracer physiology, and whole-body 3D contextual information is required rather than local-region interpretation. To advance PETRG, we propose PETRG-3D, an end-to-end 3D dual-branch framework that separately encodes PET and CT volumes and incorporates style-adaptive prompts to mitigate inter-hospital variability in reporting practices. We construct PETRG-Lym, a multi-center lymphoma dataset collected from four hospitals (824 reports w/ 245,509 paired PET/CT slices), and construct AutoPET-RG-Lym, a publicly accessible PETRG benchmark derived from open imaging data but equipped with new expert-written, clinically validated reports (135 cases). To assess clinical utility, we introduce PETRG-Score, a lymphoma-specific evaluation protocol that jointly measures metabolic and structural findings across curated anatomical regions. Experiments show that PETRG-3D substantially outperforms existing methods on both natural language metrics (e.g., +31.49% ROUGE-L) and clinical efficacy metrics (e.g., +8.18% PET-All), highlighting the benefits of volumetric dual-modality modeling and style-aware prompting. Overall, this work establishes a foundation for future PET/CT-specific models emphasizing disease-aware reasoning and clinically reliable evaluation. Codes, models, and AutoPET-RG-Lym will be released.

Accurate segmentation of tubular structures in medical images, such as vessels and airway trees, is crucial for computer-aided diagnosis, radiotherapy, and surgical planning. However, significant challenges exist in algorithm design when faced with diverse sizes, complex topologies, and (often) incomplete data annotation of these structures. We address these difficulties by proposing a new tubular structure segmentation framework named HarmonySeg. First, we design a deep-to-shallow decoder network featuring flexible convolution blocks with varying receptive fields, which enables the model to effectively adapt to tubular structures of different scales. Second, to highlight potential anatomical regions and improve the recall of small tubular structures, we incorporate vesselness maps as auxiliary information. These maps are aligned with image features through a shallow-and-deep fusion module, which simultaneously eliminates unreasonable candidates to maintain high precision. Finally, we introduce a topology-preserving loss function that leverages contextual and shape priors to balance the growth and suppression of tubular structures, which also allows the model to handle low-quality and incomplete annotations. Extensive quantitative experiments are conducted on four public datasets. The results show that our model can accurately segment 2D and 3D tubular structures and outperform existing state-of-the-art methods. External validation on a private dataset also demonstrates good generalizability.