The present study performs a comprehensive fairness analysis of machine learning (ML) models for the diagnosis of Mild Cognitive Impairment (MCI) and Alzheimer's disease (AD) from MRI-derived neuroimaging features. Biases associated with age, race, and gender in a multi-cohort dataset, as well as the influence of proxy features encoding these sensitive attributes, are investigated. The reliability of various fairness definitions and metrics in the identification of such biases is also assessed. Based on the most appropriate fairness measures, a comparative analysis of widely used pre-processing, in-processing, and post-processing bias mitigation strategies is performed. Moreover, a novel composite measure is introduced to quantify the trade-off between fairness and performance by considering the F1-score and the equalized odds ratio, making it appropriate for medical diagnostic applications. The obtained results reveal the existence of biases related to age and race, while no significant gender bias is observed. The deployed mitigation strategies yield varying improvements in terms of fairness across the different sensitive attributes and studied subproblems. For race and gender, Reject Option Classification improves equalized odds by 46% and 57%, respectively, and achieves harmonic mean scores of 0.75 and 0.80 in the MCI versus AD subproblem, whereas for age, in the same subproblem, adversarial debiasing yields the highest equalized odds improvement of 40% with a harmonic mean score of 0.69. Insights are provided into how variations in AD neuropathology and risk factors, associated with demographic characteristics, influence model fairness.

We propose a self-supervised feature learning assisted reconstruction (SSFL-Recon) framework for MRI reconstruction to address the limitation of existing supervised learning methods. Although recent deep learning-based methods have shown promising performance in MRI reconstruction, most require fully-sampled images for supervised learning, which is challenging in practice considering long acquisition times under respiratory or organ motion. Moreover, nearly all fully-sampled datasets are obtained from conventional reconstruction of mildly accelerated datasets, thus potentially biasing the achievable performance. The numerous undersampled datasets with different accelerations in clinical practice, hence, remain underutilized. To address these issues, we first train a self-supervised feature extractor on undersampled images to learn sampling-insensitive features. The pre-learned features are subsequently embedded in the self-supervised reconstruction network to assist in removing artifacts. Experiments were conducted retrospectively on an in-house 2D cardiac Cine dataset, including 91 cardiovascular patients and 38 healthy subjects. The results demonstrate that the proposed SSFL-Recon framework outperforms existing self-supervised MRI reconstruction methods and even exhibits comparable or better performance to supervised learning up to $16\times$ retrospective undersampling. The feature learning strategy can effectively extract global representations, which have proven beneficial in removing artifacts and increasing generalization ability during reconstruction.

Artificial intelligence (AI)-based chest X-ray (CXR) interpretation assistants have demonstrated significant progress and are increasingly being applied in clinical settings. However, contemporary medical AI models often adhere to a simplistic input-to-output paradigm, directly processing an image and an instruction to generate a result, where the instructions may be integral to the model's architecture. This approach overlooks the modeling of the inherent diagnostic reasoning in chest X-ray interpretation. Such reasoning is typically sequential, where each interpretive stage considers the images, the current task, and the contextual information from previous stages. This oversight leads to several shortcomings, including misalignment with clinical scenarios, contextless reasoning, and untraceable errors. To fill this gap, we construct CXRTrek, a new multi-stage visual question answering (VQA) dataset for CXR interpretation. The dataset is designed to explicitly simulate the diagnostic reasoning process employed by radiologists in real-world clinical settings for the first time. CXRTrek covers 8 sequential diagnostic stages, comprising 428,966 samples and over 11 million question-answer (Q&A) pairs, with an average of 26.29 Q&A pairs per sample. Building on the CXRTrek dataset, we propose a new vision-language large model (VLLM), CXRTrekNet, specifically designed to incorporate the clinical reasoning flow into the VLLM framework. CXRTrekNet effectively models the dependencies between diagnostic stages and captures reasoning patterns within the radiological context. Trained on our dataset, the model consistently outperforms existing medical VLLMs on the CXRTrek benchmarks and demonstrates superior generalization across multiple tasks on five diverse external datasets. The dataset and model can be found in our repository (https://github.com/guanjinquan/CXRTrek).

Deep models, such as convolutional neural networks (CNNs) and vision transformer (ViT), demonstrate remarkable performance in image classification. However, those deep models require large data to fine-tune, which is impractical in the medical domain due to the data privacy issue. Furthermore, despite the feasible performance of contrastive language image pre-training (CLIP) in the natural domain, the potential of CLIP has not been fully investigated in the medical field. To face these challenges, we considered three scenarios: 1) we introduce a novel CLIP variant using four CNNs and eight ViTs as image encoders for the classification of brain cancer and skin cancer, 2) we combine 12 deep models with two federated learning techniques to protect data privacy, and 3) we involve traditional machine learning (ML) methods to improve the generalization ability of those deep models in unseen domain data. The experimental results indicate that maxvit shows the highest averaged (AVG) test metrics (AVG = 87.03\%) in HAM10000 dataset with multimodal learning, while convnext\_l demonstrates remarkable test with an F1-score of 83.98\% compared to swin\_b with 81.33\% in FL model. Furthermore, the use of support vector machine (SVM) can improve the overall test metrics with AVG of $\sim 2\%$ for swin transformer series in ISIC2018. Our codes are available at https://github.com/AIPMLab/SkinCancerSimulation.

This study presents a multimodal AI framework designed for precisely classifying medical diagnostic images. Utilizing publicly available datasets, the proposed system compares the strengths of convolutional neural networks (CNNs) and different large language models (LLMs). This in-depth comparative analysis highlights key differences in diagnostic performance, execution efficiency, and environmental impacts. Model evaluation was based on accuracy, F1-score, average execution time, average energy consumption, and estimated $CO_2$ emission. The findings indicate that although CNN-based models can outperform various multimodal techniques that incorporate both images and contextual information, applying additional filtering on top of LLMs can lead to substantial performance gains. These findings highlight the transformative potential of multimodal AI systems to enhance the reliability, efficiency, and scalability of medical diagnostics in clinical settings.

Multimodal Large Language Models (MLLMs) are of great application across many domains, such as multimodal understanding and generation. With the development of diffusion models (DM) and unified MLLMs, the performance of image generation has been significantly improved, however, the study of image screening is rare and its performance with MLLMs is unsatisfactory due to the lack of data and the week image aesthetic reasoning ability in MLLMs. In this work, we propose a complete solution to address these problems in terms of data and methodology. For data, we collect a comprehensive medical image screening dataset with 1500+ samples, each sample consists of a medical image, four generated images, and a multiple-choice answer. The dataset evaluates the aesthetic reasoning ability under four aspects: \textit{(1) Appearance Deformation, (2) Principles of Physical Lighting and Shadow, (3) Placement Layout, (4) Extension Rationality}. For methodology, we utilize long chains of thought (CoT) and Group Relative Policy Optimization with Dynamic Proportional Accuracy reward, called DPA-GRPO, to enhance the image aesthetic reasoning ability of MLLMs. Our experimental results reveal that even state-of-the-art closed-source MLLMs, such as GPT-4o and Qwen-VL-Max, exhibit performance akin to random guessing in image aesthetic reasoning. In contrast, by leveraging the reinforcement learning approach, we are able to surpass the score of both large-scale models and leading closed-source models using a much smaller model. We hope our attempt on medical image screening will serve as a regular configuration in image aesthetic reasoning in the future.

The aim of this study was to evaluate image quality for contrast-enhanced (CE) neck MRI with a deep learning-reconstructed VIBE sequence with acceleration factors (AF) 4 (DL4-VIBE) and 6 (DL6-VIBE). Patients referred for neck MRI were examined in a 3-Tesla scanner in this prospective, single-center study. Four CE fat-saturated (FS) VIBE sequences were acquired in each patient: Star-VIBE (4:01 min), VIBE (2:05 min), DL4-VIBE (0:24 min), DL6-VIBE (0:17 min). Image quality was evaluated by three radiologists with a 5-point Likert scale and included overall image quality, muscle contour delineation, conspicuity of mucosa and pharyngeal musculature, FS uniformity, and motion artifacts. Objective image quality was assessed with signal-to-noise ratio (SNR), contrast-to-noise ratio (CNR), and quantification of metal artifacts. 68 patients (60.3% male; mean age 57.4±16 years) were included in this study. DL4-VIBE was superior for overall image quality, delineation of muscle contours, differentiation of mucosa and pharyngeal musculature, vascular delineation, and motion artifacts. Notably, DL4-VIBE exhibited exceptional FS uniformity (p<0.001). SNR and CNR were superior for DL4-VIBE compared to all other sequences (p<0.001). Metal artifacts were least pronounced in the standard VIBE, followed by DL4-VIBE (p<0.001). Although DL6-VIBE was inferior to DL4-VIBE, it demonstrated improved FS homogeneity, delineation of pharyngeal mucosa, and CNR compared to Star-VIBE and VIBE. DL4-VIBE significantly improves image quality for CE neck MRI with a fraction of the scan time of conventional sequences.

Low-grade gliomas (LGGs) are among the most problematic brain tumors to reliably segment in FLAIR MRI, and effective delineation of these lesions is critical for clinical diagnosis, treatment planning, and patient monitoring. Nevertheless, conventional U-Net-based approaches usually suffer from the loss of critical structural details owing to repetitive down-sampling, while the encoder features often retain irrelevant information that is not properly utilized by the decoder. To solve these challenges, this paper offers a dual-attention U-shaped design, named ECASE-Unet, which seamlessly integrates Efficient Channel Attention (ECA) and Squeeze-and-Excitation (SE) blocks in both the encoder and decoder stages. By selectively recalibrating channel-wise information, the model increases diagnostically significant regions of interest and reduces noise. Furthermore, dilated convolutions are introduced at the bottleneck layer to capture multi-scale contextual cues without inflating computational complexity, and dropout regularization is systematically applied to prevent overfitting on heterogeneous data. Experimental results on the Kaggle Low-Grade-Glioma dataset suggest that ECASE-Unet greatly outperforms previous segmentation algorithms, reaching a Dice coefficient of 0.9197 and an Intersection over Union (IoU) of 0.8521. Comprehensive ablation studies further reveal that integrating ECA and SE modules delivers complementing benefits, supporting the model's robust efficacy in precisely identifying LGG boundaries. These findings underline the potential of ECASE-Unet to expedite clinical operations and improve patient outcomes. Future work will focus on improving the model's applicability to new MRI modalities and studying the integration of clinical characteristics for a more comprehensive characterization of brain tumors.

This study aims to assess the added value of utilizing short-dynamic whole-body PET/CT scans and implementing motion correction before quantifying metabolic rate, offering more insights into physiological processes. While this approach may not be commonly adopted, addressing motion effects is crucial due to their demonstrated potential to cause significant errors in parametric imaging. A 15-minute dynamic FDG PET acquisition protocol was utilized for four lymphoma patients undergoing therapy evaluation. Parametric imaging was obtained using a population-based input function (PBIF) derived from twelve patients with full 65-minute dynamic FDG PET acquisition. AI-based registration methods were employed to correct misalignments between both PET and ACCT and PET-to-PET. Tumour characteristics were assessed using both parametric images and standardized uptake values (SUV). The motion correction process significantly reduced mismatches between images without significantly altering voxel intensity values, except for SUV_max. Following the alignment of the attenuation correction map with the PET frame, an increase in SUV_max in FDG-avid lymph nodes was observed, indicating its susceptibility to spatial misalignments. In contrast, Patlak K_i parameter was highly sensitive to misalignment across PET frames, that notably altered the Patlak slope. Upon completion of the motion correction process, the parametric representation revealed heterogeneous behaviour among lymph nodes compared to SUV images. Notably, reduced volume of elevated metabolic rate was determined in the mediastinal lymph nodes in contrast with an SUV of 5 g/ml, indicating potential perfusion or inflammation. Motion resolved short-dynamic PET can enhance the utility and reliability of parametric imaging, an aspect often overlooked in commercial software.

Discrete wavelet transforms have been applied in many machine learning models for the analysis of COVID-19; however, little is known about the impact of combined multilevel wavelet decompositions for the disease identification. This study proposes a computer-aided diagnosis system for addressing the combined multilevel effects of multiscale radiomic features on multiclass COVID-19 classification using chest X-ray images. A two-level discrete wavelet transform was applied to an optimal region of interest to obtain multiscale decompositions. Both approximation and detail coefficients were extensively investigated in varying frequency bands through 1240 experimental models. High dimensionality in the feature space was managed using a proposed filter- and wrapper-based feature selection approach. A comprehensive comparison was conducted between the bands and features to explore best-performing ensemble algorithm models. The results indicated that incorporating multilevel decompositions could lead to improved model performance. An inclusive region of interest, encompassing both lungs and the mediastinal regions, was identified to enhance feature representation. The light gradient-boosting machine, applied on combined bands with the features of basic, gray-level, Gabor, histogram of oriented gradients and local binary patterns, achieved the highest weighted precision, sensitivity, specificity, and accuracy of 97.50 %, 97.50 %, 98.75 %, and 97.50 %, respectively. The COVID-19-versus-the-rest receiver operating characteristic area under the curve was 0.9979. These results underscore the potential of combining decomposition levels with the original signals and employing an inclusive region of interest for effective COVID-19 detection, while the feature selection and training processes remain efficient within a practical computational time.