Proceedings of SPIE--the International Society for Optical Engineering最新文献

Emerging deep-learning-based CT denoising techniques have the potential to improve diagnostic image quality in low-dose CT exams. However, aggressive radiation dose reduction and the intrinsic uncertainty in convolutional neural network (CNN) outputs are detrimental to detecting critical lesions (e.g., liver metastases) in CNN-denoised images. To tackle these issues, we characterized CNN output distribution via total uncertainty (i.e., data + model uncertainties) and predictive mean. Local mean-uncertainty-ratio (MUR) was calculated to detect highly unreliable regions in the denoised images. A MUR-driven adaptive local fusion (ALF) process was developed to adaptively merge local predictive means with the original noisy images, thereby improving image robustness. This process was incorporated into a previously validated deep-learning model observer to quantify liver metastasis detectability, using area under localization receiver operating characteristic curve (LAUC) as the figure-of-merit. For proof-of-concept, the proposed method was established and validated for a ResNet-based CT denoising method. A recent patient abdominal CT dataset was used in validation, involving 3 lesion sizes (7, 9, and 11 mm), 3 lesion contrasts (15, 20, and 25 HU), and 3 dose levels (25%, 50%, and 100% dose). Visual inspection and quantitative analyses were conducted. Statistical significance was tested. Total uncertainty at lesions and liver background generally increased as radiation dose decreased. With fixed dose, lesion-wise MUR showed no dependency on lesion size or contrast, but exhibited large variance across lesion locations (MUR range ~0.7 to 19). Compared to original ResNet-based denoising, the MUR-driven ALF consistently improved lesion detectability in challenging conditions such as lower dose, smaller lesion size, or lower contrast (range of absolute gain in LAUC: 0.04 to 0.1; P-value 0.008). The proposed method has the potential to improve reliability of deep-learning CT denoising and enhance lesion detection.

The Text to Medical Image (T2MedI) approach using latent diffusion models holds significant promise for addressing the scarcity of medical imaging data and elucidating the appearance distribution of lesions corresponding to specific patient status descriptions. Like natural image synthesis models, our investigations reveal that the T2MedI model may exhibit biases towards certain subgroups, potentially neglecting minority groups present in the training dataset. In this study, we initially developed a T2MedI model adapted from the pre-trained Imagen framework. This model employs a fixed Contrastive Language-Image Pre-training (CLIP) text encoder, with its decoder fine-tuned using medical images from the Radiology Objects in Context (ROCO) dataset. We conduct both qualitative and quantitative analyses to examine its gender bias. To address this issue, we propose a subgroup distribution alignment method during fine-tuning on a target application dataset. Specifically, this process involves an alignment loss, guided by an off-the-shelf sensitivity-subgroup classifier, which aims to synchronize the classification probabilities between the generated images and those expected in the target dataset. Additionally, we preserve image quality through a CLIP-consistency regularization term, based on a knowledge distillation framework. For evaluation purposes, we designated the BraTS18 dataset as the target, and developed a gender classifier based on brain magnetic resonance (MR) imaging slices derived from it. Our methodology significantly mitigates gender representation inconsistencies in the generated MR images, aligning them more closely with the gender distribution in the BraTS18 dataset.

Purpose: We developed a generative model capable of producing synthetic trabecular bone that can be precisely tuned to achieve specific structural characteristics, such as bone volume fraction (BV/TV), trabecular thickness (Tb.Th), and spacing (Tb.Sp).

Methods: The generative model is based on Diffusion Transformers (DiT), a latent diffusion approach employing a transformer architecture in the denoising network. To control the microstructure characteristics of the synthetic trabecular bone samples, the model is conditioned on BV/TV, Tb.Th, and Tb.Sp. The training data involved 29898 256×256-pixel Regions of Interest (ROIs) extracted from micro-CT volumes ( $50\mu \text{m}$ voxel size) of 20 femoral bone specimens, paired with trabecular metrics computed within each ROI; the training/validation split was 9:1. For testing, 3499 synthetic bone samples were generated over a wide range of condition (target) microstructure metrics. Results were evaluated in terms of (i) the ability to cover real-world distribution of trabecular structures (coverage), (ii) agreement with target metric values (Pearson Correlation), and (iii) consistency of the metrics across multiple realizations of the DiT model with fixed condition (Coefficient of Variation, CV).

Results: The model achieved good coverage of real-world bone microstructures and visual similarity to true trabecular ROIs. Pearson Correlations against the condition (target) metric values were high: 0.9540 for BV/TV, 0.9618 for Tb.Th, and 0.9835 Tb.Sp. Microstructural characteristics of the synthetic samples were stable across DiT realizations, with CV ranging from 3.37% to 11.78% for BV/TV, 2.27% to 3.22% for Tb.Th, and 2.53% to 5.00% for Tb.Sp.

Conclusion: The proposed generative model is capable of generating realistic digital trabecular bones that can be precisely tuned to achieve specified microstructural characteristics. Possible applications include virtual clinical trials of new skeletal image biomarkers and establishing priors for advanced image reconstruction.

Protocol optimization is critical in Computed Tomography (CT) for achieving desired diagnostic image quality while minimizing radiation dose. Due to the inter-effect of influencing CT parameters, traditional optimization methods rely on the testing of exhaustive combinations of these parameters. This poses a notable limitation due to the impracticality of exhaustive parameter testing. This study introduces a novel methodology leveraging Virtual Imaging Trials (VITs) and reinforcement learning to more efficiently optimize CT protocols. Computational phantoms with liver lesions were imaged using a validated CT simulator and reconstructed with a novel CT reconstruction Toolkit. The optimization parameter space included tube voltage, tube current, reconstruction kernel, slice thickness, and pixel size. The optimization process was done using a Proximal Policy Optimization (PPO) agent which was trained to maximize the Detectability Index (d') of the liver lesion for each reconstructed image. Results showed that our reinforcement learning approach found the absolute maximum d' across the test cases while requiring 79.7% fewer steps compared to an exhaustive search, demonstrating both accuracy and computational efficiency, offering a efficient and robust framework for CT protocol optimization. The flexibility of the proposed technique allows for use of varying image quality metrics as the objective metric to maximize for. Our findings highlight the advantages of combining VIT and reinforcement learning for CT protocol management.

The increased development and production of Computed Tomography (CT) scanner technology has expanded patient access to advanced and affordable medical imaging technologies but has also introduced sources of variability in the clinical imaging landscape, which may influence patient care. This study examines the impact of intra-scanner and inter-scanner variability on image quality and quantitative imaging tasks, with a focus on the detectability index (d') as a measure of patient-specific task performance. We evaluated 813 clinical phantom image sets from the COPDGene study, aggregated by CT scanner make, model, and acquisition and reconstruction protocol. Each phantom image set was assessed for image quality metrics, including the Noise Power Spectrum (NPS) and in-plane Modulation Transfer Function (MTF). The d' index was calculated for 12 hypothetical lesion detection tasks, emulating clinically relevant lung and liver lesions of varying sizes and contrast levels. Qualitatively, analysis showed intra-scanner variability in NPS and MTF curves measured for identical acquisition and reconstruction settings. Inter-scanner comparisons demonstrated variability in d' measurements across different scanner makes and models, of similar acquisition and reconstruction settings. The study showed an intra-scanner variability of up to 13.7% and an inter-scanner variability of up to 19.3% in the d' index. These findings emphasize the need for considering scanner variability in patient-centered care and indicate that CT technology may influence the reliability of imaging tasks. The results of this study further motivate the development of virtual scanner models to better model and mitigate the variability observed in the clinical imaging landscape.

Recently, the use of circle representation has emerged as a method to improve the identification of spherical objects (such as glomeruli, cells, and nuclei) in medical imaging studies. In traditional bounding box-based object detection, combining results from multiple models improves accuracy, especially when real-time processing isn't crucial. Unfortunately, this widely adopted strategy is not readily available for combining circle representations. In this paper, we propose Weighted Circle Fusion (WCF), a simple approach for merging predictions from various circle detection models. Our method leverages confidence scores associated with each proposed bounding circle to generate averaged circles. We evaluate our method on a proprietary dataset for glomerular detection in whole slide imaging (WSI) and find a performance gain of 5% compared to existing ensemble methods. Additionally, we assess the efficiency of two annotation methods-fully manual annotation and a human-in-the-loop (HITL) approach-in labeling 200,000 glomeruli. The HITL approach, which integrates machine learning detection with human verification, demonstrated remarkable improvements in annotation efficiency. The Weighted Circle Fusion technique not only enhances object detection precision but also notably reduces false detections, presenting a promising direction for future research and application in pathological image analysis. The source code has been made publicly available at https://github.com/hrlblab/WeightedCircleFusion.

Purpose: Bankart lesions, or anterior-inferior glenoid labral tears, are diagnostically challenging on standard MRIs due to their subtle imaging features-often necessitating invasive MRI arthrograms (MRAs). This study develops deep learning (DL) models to detect Bankart lesions on both standard MRIs and MRAs, aiming to improve diagnostic accuracy and reduce reliance on MRAs.

Methods: We curated a dataset of 586 shoulder MRIs (335 standard, 251 MRAs) from 558 patients who underwent arthroscopy. Ground truth labels were derived from intraoperative findings, the gold standard for Bankart lesion diagnosis. Separate DL models for MRAs and standard MRIs were trained using the Swin Transformer architecture, pre-trained on a public knee MRI dataset. Predictions from sagittal, axial, and coronal views were ensembled to optimize performance. The models were evaluated on a 20% hold-out test set (117 MRIs: 46 MRAs, 71 standard MRIs).

Results: Bankart lesions were identified in 31.9% of MRAs and 8.6% of standard MRIs. The models achieved AUCs of 0.87 (86% accuracy, 83% sensitivity, 86% specificity) and 0.90 (85% accuracy, 82% sensitivity, 86% specificity) on standard MRIs and MRAs, respectively. These results match or surpass radiologist performance on our dataset and reported literature metrics. Notably, our model's performance on non-invasive standard MRIs matched or surpassed the radiologists interpreting MRAs.

Conclusion: This study demonstrates the feasibility of using DL to address the diagnostic challenges posed by subtle pathologies like Bankart lesions. Our models demonstrate potential to improve diagnostic confidence, reduce reliance on invasive imaging, and enhance accessibility to care.

Low-dose CT simulation is needed to assess reconstruction/denoising techniques and optimize dose. Projection-domain noise-insertion methods require manufacturers' proprietary tools. Image-domain noise-insertion methods face various challenges that affect generalizability, and few have been systematically validated for 3D noise synthesis. To improve generalizability, we presented a physics-informed model-based generative neural network for simulating scanner- and algorithm-specific low-dose CT exams (PALETTE). PALETTE included a noise-prior-generation process, a Noise2Noisier sub-network, and a noise-texture-synthesis sub-network. Custom regularization terms were developed to enforce 3D noise texture quality. Using PALETTE, one 2D and two 2.5D models (denoted as 2.5D N-N and N-1) were developed to conduct 2D and effective 3D noise modeling, respectively (input/output images: 2D - 1/1, 2.5D N-N - 3/3, 2.5D N-1 - 5/1). These models were trained and tested with an open-access abdominal CT dataset, including 20 testing cases reconstructed with two kernels and various field-of-view. In visual inspection, the 2D and 2.5D N-N models generated realistic local and global noise texture, while 2.5D N-1 showed more perceptual difference using the sharper kernel and coronal reformat. In quantitative evaluation, local noise level was compared using mean-absolute-percent-difference (MAPD), and global spectral similarity was assessed using spectral correlation mapper (SCM) and spectral angle mapper (SAM). The 2D model provided equivalent or relatively better performance than 2.5D models, showing well-matched local noise levels and high spectral similarity compared to the reference (sharper/smoother kernels): MAPD - 2D 1.5%/5.6% (p>0.05), 2.5D N-N 8.5%/7.9% (p<0.05), 2.5D N-1 12.3%/10.9% (p<0.05); mean SCM - 2D 0.97/0.97, 2.5D N-N 0.96/0.97, 2.5D N-1 0.85/0.97; mean SAM - 2D 0.12/0.12, 2.5D N-N 0.14/0.12, 2.5D N-1 0.37/0.12. With tripled model width, the 2.5D N-N outperformed N-1. This indicated 2.5D models need more learning capacity to further enhance 3D noise modeling. Using physics-based prior information, PALETTE can provide high-quality low-dose CT simulation to resemble scanner- and algorithm-specific 3D noise characteristics.

Task-based image quality assessment is essential for CT protocol and radiation dose optimization. Despite many ongoing efforts, there is still an unmet need to measure and monitor the quality of images acquired from each patient exam. In this work, we developed a patient-specific channelized Hotelling observer (CHO)-based method to estimate the lesion detectability for each individual patient scan. The ensemble of background was created from patient images to include both relatively uniform regions and anatomically varying regions. Signals were modelled from lesions of different sizes and contrast levels after incorporating the effect of contrast-dependent spatial resolution. Index of detectability (d') was estimated using a CHO framework. This method was applied to clinical patient images obtained from a CT scanner at 3 different radiation dose levels. The d' for 5 different lesion size/contrast conditions was calculated across the scan range of each patient exam. The average noise levels and the d' averaged from 5 conditions were 13.2/3.78, 17.1/2.93 and 21.9/2.43 at 100%, 50% and 25% dose levels, respectively.

Chest radiographs are a vital tool for identifying pathological changes within the thoracic cavity. Artificial intelligence (AI) and machine learning (ML) driven screening or diagnostic applications require accurate detection of anatomical structures within the Chest X-ray (CXR) image. The You Only Look Once (YOLO) object detection models have recently gained prominence for their efficacy in detecting anatomical structures in medical images. However, state-of-the-art results using it are typically for single anatomical organ detection. Advanced image analysis would benefit from simultaneous detection more than one anatomical organ. In this work we propose a multi-organ detection technique through two recent YOLO versions and their sub-variants. We evaluate their effectiveness in detecting lung and heart regions in CXRs simultaneously. We used the JSRT CXR dataset for internal training, validation, and testing. Further, the generalizability of the models is evaluated using two external test sets, viz., the Montgomery CXR dataset and a subset of the RSNA CXR dataset against available annotations therein. Our evaluation demonstrates that YOLOv9 models notably outperform YOLOv8 variants. We demonstrated further improvements in detection performance through ensemble approaches.