The deployment of Large Language Models (LLMs) for structuring clinical data is critically hindered by their tendency to hallucinate facts and their inability to follow domain-specific rules. To address this, we introduce MedPAO, a novel agentic framework that ensures accuracy and verifiable reasoning by grounding its operation in established clinical protocols such as the ABCDEF protocol for CXR analysis. MedPAO decomposes the report structuring task into a transparent process managed by a Plan-Act-Observe (PAO) loop and specialized tools. This protocol-driven method provides a verifiable alternative to opaque, monolithic models. The efficacy of our approach is demonstrated through rigorous evaluation: MedPAO achieves an F1-score of 0.96 on the critical sub-task of concept categorization. Notably, expert radiologists and clinicians rated the final structured outputs with an average score of 4.52 out of 5, indicating a level of reliability that surpasses baseline approaches relying solely on LLM-based foundation models. The code is available at: https://github.com/MiRL-IITM/medpao-agent

Functional MRI currently supports a limited application space stemming from modest dataset sizes, large interindividual variability and heterogeneity among scanning protocols. These constraints have made it difficult for fMRI researchers to take advantage of modern deep-learning tools that have revolutionized other fields such as NLP, speech transcription, and image recognition. To address these issues, we scaled up functional connectome fingerprinting as a neural network pre-training task, drawing inspiration from speaker recognition research, to learn a generalizable representation of brain function. This approach sets a new high-water mark for neural fingerprinting on a previously unseen scale, across many popular public fMRI datasets (individual recognition over held out scan sessions: 94% on MPI-Leipzig, 94% on NKI-Rockland, 73% on OASIS-3, and 99% on HCP). Near-ceiling performance is maintained even when the duration of the evaluation scan is truncated to less than two minutes. We show that this representation can also generalize to support accurate neural fingerprinting for completely new datasets and participants not used in training. Finally, we demonstrate that the representation learned by the network encodes features related to individual variability that supports some transfer learning to new tasks. These results open the door for a new generation of clinical applications based on functional imaging data.

Lung cancer is a leading cause of cancer-related mortality worldwide. Accurate staging of mediastinal lymph nodes is a crucial step in determining appropriate treatment approaches. Current noninvasive diagnostic methods do not provide sufficient accuracy to confidently decide on surgery without histological confirmation. Our study aimed to develop a artificial intelligence model for the precise prediction of N2 lymph node metastasis. We retrospectively analyzed 1489 patients who underwent standard cervical mediastinoscopy at our department, including 472 patients diagnosed with non-small cell lung cancer. We developed three distinct prediction models for N2 lymph node station metastasis: one using standard statistical analysis, another utilizing an image processing deep learning algorithm with thoracic CT, and the third employing various machine learning methods with clinicopathological and radiological data. We compared diagnostic accuracy, area under the curve (AUC), sensitivity, and specificity rates, as well as the F1-score of all models. Linear discriminant analysis, quadratic discriminant analysis, Gaussian naive Bayes, and artificial neural networks all surpassed 90% accuracy. The linear support vector machine demonstrated the highest performance, with an accuracy of 95.7%, an AUC of 93.5%, and an F1-score of 92%, respectively and outperformed the logistic regression-based statistical model, which reached an accuracy of 90.6% and an AUC of 85.7%. Machine learning models outperformed standard statistical analysis models in predicting N2 lymph node metastasis. Implementing these machine learning prediction models might greatly improve the accuracy of mediastinal lymph node metastasis detection, thereby enhancing clinical decision making and patient outcomes.

Hematologic toxicity (HT) is a common and serious side effect for advanced cervical cancer patients undergoing chemoradiotherapy. Accurately predicting HT can significantly improve patient management and treatment outcomes. This study aims to develop and evaluate interpretable machine learning models that use radiomic and dosimetric features to predict HT in advanced cervical cancer patients. Retrospectively collected general clinical data, planning CT images, and dose files from 205 patients with advanced cervical cancer who underwent chemoradiotherapy, and classified them according to the severity of HT. Radiomics and dosiomics features were extracted from the same region of interest, and feature selection was performed using a random forest algorithm. Radiomics models, dosiomics models, and hybrid models were then constructed based on extreme gradient boosting trees. Sensitivity, specificity, and the area under the receiver operating characteristic curve (AUC) were calculated to evaluate the classification performance of the models. Finally, SHAP values were used to perform interpretability analysis on the best model to enhance the transparency and practicality of the model. The sensitivity, specificity, and AUC values for the radiomics model were 0.42, 0.86, and 0.78, respectively, while those for the dosiomics model were 0.50, 0.90, and 0.74. In contrast, the hybrid model exhibited superior classification performance with sensitivity, specificity, and AUC values of 0.50, 0.83, and 0.83, respectively. Compared to the standalone radiomics and dosiomics models, the hybrid model demonstrated enhanced classification capability. Interpretability analysis based on SHAP values not only provided a ranking of feature importance and the distribution of feature impacts on model outputs but also elucidated the specific decision-making processes influenced by these features and the interactions between them. This enables clinicians to gain a more intuitive understanding of the model's decisions. For patients with advanced cervical cancer undergoing chemoradiotherapy, the integration of radiomics and dosiomics features can significantly enhance the classification performance of predictive models, thereby holding considerable potential for refining patient treatment strategies. Interpretability analysis based on SHAP values can aid clinicians in more readily understanding the model's decisions, thus promoting the effective implementation of the model in clinical practice.

Brain tumour identification, segmentation cataloguing from MRI images is most thought-provoking and is a very much essential for many medical image analysis applications. Every brain imaging modality provides information about various parts of the tumor. In current years deep learning systems have shown auspicious outcomes in medical image investigation tasks. Despite several recent works achieved a significant result on brain tumour segmentation and classification, they come with an improved performance at the expense of increased computational complexity to train and test the system. This exploration paper investigates the efficacy of popular deep learning architectures namely Xception Net, MobileNet for classification and DeepLab for segmentation of the cancerous region of brain tumor. Each architecture is trained using a BRATS 2018 dataset and evaluated for its performance in accurately classifying tumor presence and delineating tumor boundaries. The DeepLab models accomplished a best segmentation result with Pearson Correlation Coefficient values 0.50 respectively and the deep learning models Xception Net and MobileNet achieved an accuracy of 0.8921 and 0.9176respectively. The experimental results show that these architectures achieve high accuracy and precise segmentation. The findings of this study contribute to advancing the field of medical image analysis and hold implications for improving the analysis and dealing of brain tumors.

Frontotemporal dementia (FTD) presents a complex spectrum of neurodegenerative disorders, encompassing distinct subtypes with varied clinical manifestations. This study investigates alterations in brain module organization associated with FTD subtypes using connectome analysis, aiming to identify potential biomarkers and enhance subtype prediction. Resting-state functional magnetic resonance imaging data were obtained from 41 individuals with behavioral variant frontotemporal dementia (BV-FTD), 32 with semantic variant frontotemporal dementia (SV-FTD), 28 with progressive non-fluent aphasia frontotemporal dementia (PNFA-FTD), and 94 healthy controls. Individual functional brain networks were constructed at the voxel level and binarized based on density thresholds. Modular segregation index (MSI) and participation coefficient (PC) were calculated to assess module integrity and identify regions with altered nodal properties. The relationship between modular measures and clinical scores was examined, and machine learning models were developed for subtype prediction. Both BV-FTD and SV-FTD groups exhibited decreased MSI in the subcortical module (SUB), default mode network (DMN), and ventral attention network (VAN) compared to healthy controls. Additionally, BV-FTD specifically displayed disrupted frontoparietal network (FPN) integrity compared to other FTD subtypes and controls. All FTD subtypes showed increased PC values in the insular region and reduced connections between the insular and VAN/FPN compared to controls. Moreover, significant associations between specific network alterations and clinical variables were observed. Machine learning models utilizing these matrices achieved high performance in differentiating FTD subtypes. This pilot study reveals diverse brain module organization across FTD subtypes, shedding light on both shared and distinct neurobiological underpinnings of the disorder.

Shared decision-making (SDM) is a cornerstone of patient-centred care, yet it has been underused in radiology. To translate research into innovative strategies to empower radiology leaders to apply SDM and outline the cultural and structural changes required for meaningful integration into clinical practice. This article synthesises case examples and evidence across imaging scenarios, evaluates emerging innovations and highlights leadership levers that can embed SDM as a core practice in radiology. Leadership interventions can transform radiology's contribution to SDM. Cases such as incidental pulmonary nodules, breast MRI in familial risk and Li-Fraumeni syndrome illustrate how radiologists can engage directly in preference-sensitive decisions. Key strategies include improving access to imaging data, using patient-friendly summaries, expanding opportunities for direct communication and incorporating patient-reported outcome measures, patient-reported experience measures and artificial intelligence (AI)-driven tools to support patient understanding. Barriers such as workflow demands, medicolegal uncertainty and lack of incentives can be addressed through leadership-driven reforms. Radiology plays a central role in care pathways, offers clinical and technical expertise and increasing patient-facing innovation. Leaders who embed SDM into training, workflows and systems can enhance radiology as a model of cutting-edge, patient-centred care. Clear actions include training, protected time, incentives, strategic application of AI and transformational leadership.

Physics-informed machine learning (PIML) is emerging as a potentially transformative paradigm for modeling complex biomedical systems by integrating parameterized physical laws with data-driven methods. Here, we review three main classes of PIML frameworks: physics-informed neural networks (PINNs), neural ordinary differential equations (NODEs), and neural operators (NOs), highlighting their growing role in biomedical science and engineering. We begin with PINNs, which embed governing equations into deep learning models and have been successfully applied to biosolid and biofluid mechanics, mechanobiology, and medical imaging among other areas. We then review NODEs, which offer continuous-time modeling, especially suited to dynamic physiological systems, pharmacokinetics, and cell signaling. Finally, we discuss deep NOs as powerful tools for learning mappings between function spaces, enabling efficient simulations across multiscale and spatially heterogeneous biological domains. Throughout, we emphasize applications where physical interpretability, data scarcity, or system complexity make conventional black-box learning insufficient. We conclude by identifying open challenges and future directions for advancing PIML in biomedical science and engineering, including issues of uncertainty quantification, generalization, and integration of PIML and large language models.

Medical imaging has revolutionized diagnosis, yet unnecessary procedures are rising, exposing patients to radiation and stress, limiting equitable access, and straining healthcare systems. The American College of Radiology Appropriateness Criteria, developed through extensive multidisciplinary review, provide evidence-based guidance but remain underutilized. Leveraging advances in LLM reasoning, we introduce a Reasoning Agent trained with Reinforcement Learning (RL), specifically Group Relative Policy Optimization (GRPO), to replicate expert clinical reasoning from the ACR Criteria. We present a novel RL approach for structured medical reasoning, systematically comparing reasoning-focused reward functions and evidence integration strategies. Our lightweight 8B model, MedReason-Embed, improves macro F1 by 18% over baseline, shows stronger reasoning alignment, and outperforms both larger and alternatively trained models, showing that reasoning-based supervision enables efficient, trustworthy clinical AI. Building on this, we design a modular end-to-end agentic architecture that automates imaging referrals: mapping diagnoses to ICD codes, retrieving PubMed evidence, and recommending optimal procedures. Crucially, the ability to generalize beyond static ACR guidelines not only enables clinicians to handle out-of-distribution cases, but also supports scaling the guideline development process itself, potentially reducing the significant effort required to create and update them. This work shows the potential of reasoning-focused RL within agentic architectures to deliver transparent, scalable, and reliable clinical decision support. Our code is available at: https://anonymous.4open.science/r/agentic-imaging-recommender-iclr-877D

Focal Cortical Dysplasia (FCD) is a primary cause of drug-resistant epilepsy and is difficult to detect in brain {magnetic resonance imaging} (MRI) due to the subtle and small-scale nature of its lesions. Accurate segmentation of FCD regions in 3D multimodal brain MRI images is essential for effective surgical planning and treatment. However, this task remains highly challenging due to the limited availability of annotated FCD datasets, the extremely small size and weak contrast of FCD lesions, the complexity of handling 3D multimodal inputs, and the need for output smoothness and anatomical consistency, which is often not addressed by standard voxel-wise loss functions. This paper presents a new framework for segmenting FCD regions in 3D brain MRI images. We adopt state-of-the-art transformer-enhanced encoder-decoder architecture and introduce a novel loss function combining Dice loss with an anisotropic {Total Variation} (TV) term. This integration encourages spatial smoothness and reduces false positive clusters without relying on post-processing. The framework is evaluated on a public FCD dataset with 85 epilepsy patients and demonstrates superior segmentation accuracy and consistency compared to standard loss formulations. The model with the proposed TV loss shows an 11.9\% improvement on the Dice coefficient and 13.3\% higher precision over the baseline model. Moreover, the number of false positive clusters is reduced by 61.6%