Interpretability can improve the safety, transparency and trust of AI models, which is especially important in healthcare applications where decisions often carry significant consequences. Mechanistic interpretability, particularly through the use of sparse autoencoders (SAEs), offers a promising approach for uncovering human-interpretable features within large transformer-based models. In this study, we apply Matryoshka-SAE to the radiology-specialised multimodal large language model, MAIRA-2, to interpret its internal representations. Using large-scale automated interpretability of the SAE features, we identify a range of clinically relevant concepts - including medical devices (e.g., line and tube placements, pacemaker presence), pathologies such as pleural effusion and cardiomegaly, longitudinal changes and textual features. We further examine the influence of these features on model behaviour through steering, demonstrating directional control over generations with mixed success. Our results reveal practical and methodological challenges, yet they offer initial insights into the internal concepts learned by MAIRA-2 - marking a step toward deeper mechanistic understanding and interpretability of a radiology-adapted multimodal large language model, and paving the way for improved model transparency. We release the trained SAEs and interpretations: https://huggingface.co/microsoft/maira-2-sae.

Medical image synthesis presents unique challenges due to the inherent complexity and high-resolution details required in clinical contexts. Traditional generative architectures such as Generative Adversarial Networks (GANs) or Variational Auto Encoder (VAEs) have shown great promise for high-resolution image generation but struggle with preserving fine-grained details that are key for accurate diagnosis. To address this issue, we introduce Pixel Perfect MegaMed, the first vision-language foundation model to synthesize images at resolutions of 1024x1024. Our method deploys a multi-scale transformer architecture designed specifically for ultra-high resolution medical image generation, enabling the preservation of both global anatomical context and local image-level details. By leveraging vision-language alignment techniques tailored to medical terminology and imaging modalities, Pixel Perfect MegaMed bridges the gap between textual descriptions and visual representations at unprecedented resolution levels. We apply our model to the CheXpert dataset and demonstrate its ability to generate clinically faithful chest X-rays from text prompts. Beyond visual quality, these high-resolution synthetic images prove valuable for downstream tasks such as classification, showing measurable performance gains when used for data augmentation, particularly in low-data regimes. Our code is accessible through the project website - https://tehraninasab.github.io/pixelperfect-megamed.

Radiologists rely on eye movements to navigate and interpret medical images. A trained radiologist possesses knowledge about the potential diseases that may be present in the images and, when searching, follows a mental checklist to locate them using their gaze. This is a key observation, yet existing models fail to capture the underlying intent behind each fixation. In this paper, we introduce a deep learning-based approach, RadGazeIntent, designed to model this behavior: having an intention to find something and actively searching for it. Our transformer-based architecture processes both the temporal and spatial dimensions of gaze data, transforming fine-grained fixation features into coarse, meaningful representations of diagnostic intent to interpret radiologists' goals. To capture the nuances of radiologists' varied intention-driven behaviors, we process existing medical eye-tracking datasets to create three intention-labeled subsets: RadSeq (Systematic Sequential Search), RadExplore (Uncertainty-driven Exploration), and RadHybrid (Hybrid Pattern). Experimental results demonstrate RadGazeIntent's ability to predict which findings radiologists are examining at specific moments, outperforming baseline methods across all intention-labeled datasets.

<i>"Just Accepted" papers have undergone full peer review and have been accepted for publication in <i>Radiology: Artificial Intelligence</i>. This article will undergo copyediting, layout, and proof review before it is published in its final version. Please note that during production of the final copyedited article, errors may be discovered which could affect the content.</i> Purpose To develop a deep learning model that uses a single inspiratory chest CT scan to generate parametric response maps (PRM) and predict functional small airway disease (fSAD). Materials and Methods In this retrospective study, predictive and generative deep learning models for PRM using inspiratory chest CT were developed using a model development dataset with fivefold cross-validation, with PRM derived from paired respiratory CT as the reference standard. Voxel-wise metrics, including sensitivity, area under the receiver operating characteristic curve (AUC), and structural similarity, were used to evaluate model performance in predicting PRM and expiratory CT images. The best performing model was tested on three internal test sets and an external test set. Results The model development dataset of 308 patients (median age, 67 years, [IQR: 62-70 years]; 113 female) was divided into the training set (<i>n</i> = 216), the internal validation set (<i>n</i> = 31), and the first internal test set (<i>n</i> = 61). The generative model outperformed the predictive model in detecting fSAD (sensitivity 86.3% vs 38.9%; AUC 0.86 vs 0.70). The generative model performed well in the second internal (AUCs of 0.64, 0.84, 0.97 for emphysema, fSAD and normal lung tissue), the third internal (AUCs of 0.63, 0.83, 0.97), and the external (AUCs of 0.58, 0.85, 0.94) test sets. Notably, the model exhibited exceptional performance in the PRISm group of the fourth internal test set (AUC = 0.62, 0.88, and 0.96). Conclusion The proposed generative model, using a single inspiratory CT, outperformed existing algorithms in PRM evaluation, achieved comparable results to paired respiratory CT. Published under a CC BY 4.0 license.

<i>"Just Accepted" papers have undergone full peer review and have been accepted for publication in <i>Radiology: Artificial Intelligence</i>. This article will undergo copyediting, layout, and proof review before it is published in its final version. Please note that during production of the final copyedited article, errors may be discovered which could affect the content.</i> Purpose To develop a collaborative AI system that integrates eye gaze data and radiology reports to improve diagnostic accuracy in chest radiograph interpretation by identifying and correcting perceptual errors. Materials and Methods This retrospective study utilized public datasets REFLACX and EGD-CXR to develop a collaborative AI solution, named Collaborative Radiology Expert (CoRaX). It employs a large multimodal model to analyze image embeddings, eye gaze data, and radiology reports, aiming to rectify perceptual errors in chest radiology. The proposed system was evaluated using two simulated error datasets featuring random and uncertain alterations of five abnormalities. Evaluation focused on the system's referral-making process, the quality of referrals, and its performance within collaborative diagnostic settings. Results In the random masking-based error dataset, 28.0% (93/332) of abnormalities were altered. The system successfully corrected 21.3% (71/332) of these errors, with 6.6% (22/332) remaining unresolved. The accuracy of the system in identifying the correct regions of interest for missed abnormalities was 63.0% [95% CI: 59.0%, 68.0%], and 85.7% (240/280) of interactions with radiologists were deemed satisfactory, meaning that the system provided diagnostic aid to radiologists. In the uncertainty-masking-based error dataset, 43.9% (146/332) of abnormalities were altered. The system corrected 34.6% (115/332) of these errors, with 9.3% (31/332) unresolved. The accuracy of predicted regions of missed abnormalities for this dataset was 58.0% [95% CI: 55.0%, 62.0%], and 78.4% (233/297) of interactions were satisfactory. Conclusion The CoRaX system can collaborate efficiently with radiologists and address perceptual errors across various abnormalities in chest radiographs. ©RSNA, 2025.

Bronchopulmonary dysplasia (BPD) is a chronic lung disease affecting 35% of extremely low birth weight infants. Defined by oxygen dependence at 36 weeks postmenstrual age, it causes lifelong respiratory complications. However, preventive interventions carry severe risks, including neurodevelopmental impairment, ventilator-induced lung injury, and systemic complications. Therefore, early BPD prognosis and prediction of BPD outcome is crucial to avoid unnecessary toxicity in low risk infants. Admission radiographs of extremely preterm infants are routinely acquired within 24h of life and could serve as a non-invasive prognostic tool. In this work, we developed and investigated a deep learning approach using chest X-rays from 163 extremely low-birth-weight infants ($\leq$32 weeks gestation, 401-999g) obtained within 24 hours of birth. We fine-tuned a ResNet-50 pretrained specifically on adult chest radiographs, employing progressive layer freezing with discriminative learning rates to prevent overfitting and evaluated a CutMix augmentation and linear probing. For moderate/severe BPD outcome prediction, our best performing model with progressive freezing, linear probing and CutMix achieved an AUROC of 0.78 $\pm$ 0.10, balanced accuracy of 0.69 $\pm$ 0.10, and an F1-score of 0.67 $\pm$ 0.11. In-domain pre-training significantly outperformed ImageNet initialization (p = 0.031) which confirms domain-specific pretraining to be important for BPD outcome prediction. Routine IRDS grades showed limited prognostic value (AUROC 0.57 $\pm$ 0.11), confirming the need of learned markers. Our approach demonstrates that domain-specific pretraining enables accurate BPD prediction from routine day-1 radiographs. Through progressive freezing and linear probing, the method remains computationally feasible for site-level implementation and future federated learning deployments.

Phrase grounding, i.e., mapping natural language phrases to specific image regions, holds significant potential for disease localization in medical imaging through clinical reports. While current state-of-the-art methods rely on discriminative, self-supervised contrastive models, we demonstrate that generative text-to-image diffusion models, leveraging cross-attention maps, can achieve superior zero-shot phrase grounding performance. Contrary to prior assumptions, we show that fine-tuning diffusion models with a frozen, domain-specific language model, such as CXR-BERT, substantially outperforms domain-agnostic counterparts. This setup achieves remarkable improvements, with mIoU scores doubling those of current discriminative methods. These findings highlight the underexplored potential of generative models for phrase grounding tasks. To further enhance performance, we introduce Bimodal Bias Merging (BBM), a novel post-processing technique that aligns text and image biases to identify regions of high certainty. BBM refines cross-attention maps, achieving even greater localization accuracy. Our results establish generative approaches as a more effective paradigm for phrase grounding in the medical imaging domain, paving the way for more robust and interpretable applications in clinical practice. The source code and model weights are available at https://github.com/Felix-012/generate_to_ground.

Foundation models have significantly revolutionized the field of chest X-ray diagnosis with their ability to transfer across various diseases and tasks. However, previous works have predominantly utilized self-supervised learning from medical image-text pairs, which falls short in dense medical prediction tasks due to their sole reliance on such coarse pair supervision, thereby limiting their applicability to detailed diagnostics. In this paper, we introduce a Dense Chest X-ray Foundation Model (DCXFM), which utilizes mixed supervision types (i.e., text, label, and segmentation masks) to significantly enhance the scalability of foundation models across various medical tasks. Our model involves two training stages: we first employ a novel self-distilled multimodal pretraining paradigm to exploit text and label supervision, along with local-to-global self-distillation and soft cross-modal contrastive alignment strategies to enhance localization capabilities. Subsequently, we introduce an efficient cost aggregation module, comprising spatial and class aggregation mechanisms, to further advance dense prediction tasks with densely annotated datasets. Comprehensive evaluations on three tasks (phrase grounding, zero-shot semantic segmentation, and zero-shot classification) demonstrate DCXFM's superior performance over other state-of-the-art medical image-text pretraining models. Remarkably, DCXFM exhibits powerful zero-shot capabilities across various datasets in phrase grounding and zero-shot semantic segmentation, underscoring its superior generalization in dense prediction tasks.

Bronchopulmonary dysplasia (BPD) is a chronic lung disease affecting 35% of extremely low birth weight infants. Defined by oxygen dependence at 36 weeks postmenstrual age, it causes lifelong respiratory complications. However, preventive interventions carry severe risks, including neurodevelopmental impairment, ventilator-induced lung injury, and systemic complications. Therefore, early BPD prognosis and prediction of BPD outcome is crucial to avoid unnecessary toxicity in low risk infants. Admission radiographs of extremely preterm infants are routinely acquired within 24h of life and could serve as a non-invasive prognostic tool. In this work, we developed and investigated a deep learning approach using chest X-rays from 163 extremely low-birth-weight infants ($\leq$32 weeks gestation, 401-999g) obtained within 24 hours of birth. We fine-tuned a ResNet-50 pretrained specifically on adult chest radiographs, employing progressive layer freezing with discriminative learning rates to prevent overfitting and evaluated a CutMix augmentation and linear probing. For moderate/severe BPD outcome prediction, our best performing model with progressive freezing, linear probing and CutMix achieved an AUROC of 0.78 $\pm$ 0.10, balanced accuracy of 0.69 $\pm$ 0.10, and an F1-score of 0.67 $\pm$ 0.11. In-domain pre-training significantly outperformed ImageNet initialization (p = 0.031) which confirms domain-specific pretraining to be important for BPD outcome prediction. Routine IRDS grades showed limited prognostic value (AUROC 0.57 $\pm$ 0.11), confirming the need of learned markers. Our approach demonstrates that domain-specific pretraining enables accurate BPD prediction from routine day-1 radiographs. Through progressive freezing and linear probing, the method remains computationally feasible for site-level implementation and future federated learning deployments.

To test the performance of a DL model developed and validated for screen-detected pulmonary nodules on incidental nodules detected in a clinical setting. A retrospective dataset of incidental pulmonary nodules sized 5-15 mm was collected, and a subset of size-matched solid nodules was selected. The performance of the DL model was compared to the Brock model. AUCs with 95% CIs were compared using the DeLong method. Sensitivity and specificity were determined at various thresholds, using a 10% threshold for the Brock model as reference. The model's calibration was visually assessed. The dataset included 49 malignant and 359 benign solid or part-solid nodules, and the size-matched dataset included 47 malignant and 47 benign solid nodules. In the complete dataset, AUCs [95% CI] were 0.89 [0.85, 0.93] for the DL model and 0.86 [0.81, 0.92] for the Brock model (p = 0.27). In the size-matched subset, AUCs of the DL and Brock models were 0.78 [0.69, 0.88] and 0.58 [0.46, 0.69] (p < 0.01), respectively. At a 10% threshold, the Brock model had a sensitivity of 0.49 [0.35, 0.63] and a specificity of 0.92 [0.89, 0.94]. At a threshold of 17%, the DL model matched the specificity of the Brock model at the 10% threshold, but had a higher sensitivity (0.57 [0.43, 0.71]). Calibration analysis revealed that the DL model overestimated the malignancy probability. The DL model demonstrated good discriminatory performance in a dataset of incidental nodules and outperformed the Brock model, but may need recalibration for clinical practice. Question What is the performance of a DL model for pulmonary nodule malignancy risk estimation developed on screening data in a dataset of incidentally detected nodules? Findings The DL model performed well on a dataset of nodules from clinical routine care and outperformed the Brock model in a size-matched subset. Clinical relevance This study provides further evidence about the potential of DL models for risk stratification of incidental nodules, which may improve nodule management in routine clinical practice.