Computerized Medical Imaging and Graphics最新文献

Accurate segmentation of the pancreas in computed tomography (CT) holds paramount importance in diagnostics, surgical planning, and interventions. Recent studies have proposed supervised deep-learning models for segmentation, but their efficacy relies on the quality and quantity of the training data. Most of such works employed small-scale public datasets, without proving the efficacy of generalization to external datasets. This study explored the optimization of pancreas segmentation accuracy by pinpointing the ideal dataset size, understanding resource implications, examining manual refinement impact, and assessing the influence of anatomical subregions. We present the AIMS-1300 dataset encompassing 1,300 CT scans. Its manual annotation by medical experts required 938 h. A 2.5D UNet was implemented to assess the impact of training sample size on segmentation accuracy by partitioning the original AIMS-1300 dataset into 11 smaller subsets of progressively increasing numerosity. The findings revealed that training sets exceeding 440 CTs did not lead to better segmentation performance. In contrast, nnU-Net and UNet with Attention Gate reached a plateau for 585 CTs. Tests on generalization on the publicly available AMOS-CT dataset confirmed this outcome. As the size of the partition of the AIMS-1300 training set increases, the number of error slices decreases, reaching a minimum with 730 and 440 CTs, for AIMS-1300 and AMOS-CT datasets, respectively. Segmentation metrics on the AIMS-1300 and AMOS-CT datasets improved more on the head than the body and tail of the pancreas as the dataset size increased. By carefully considering the task and the characteristics of the available data, researchers can develop deep learning models without sacrificing performance even with limited data. This could accelerate developing and deploying artificial intelligence tools for pancreas surgery and other surgical data science applications.

Oral squamous cell carcinoma recognition presents a challenge due to late diagnosis and costly data acquisition. A cost-efficient, computerized screening system is crucial for early disease detection, minimizing the need for expert intervention and expensive analysis. Besides, transparency is essential to align these systems with critical sector applications. Explainable Artificial Intelligence (XAI) provides techniques for understanding models. However, current XAI is mostly data-driven and focused on addressing developers’ requirements of improving models rather than clinical users’ demands for expressing relevant insights. Among different XAI strategies, we propose a solution composed of Case-Based Reasoning paradigm to provide visual output explanations and Informed Deep Learning (IDL) to integrate medical knowledge within the system. A key aspect of our solution lies in its capability to handle data imperfections, including labeling inaccuracies and artifacts, thanks to an ensemble architecture on top of the deep learning (DL) workflow. We conducted several experimental benchmarks on a dataset collected in collaboration with medical centers. Our findings reveal that employing the IDL approach yields an accuracy of 85%, surpassing the 77% accuracy achieved by DL alone. Furthermore, we measured the human-centered explainability of the two approaches and IDL generates explanations more congruent with the clinical user demands.

Spontaneous intracerebral hemorrhage (ICH) is a type of stroke less prevalent than ischemic stroke but associated with high mortality rates. Hematoma expansion (HE) is an increase in the bleeding that affects 30%–38% of hemorrhagic stroke patients. It is observed within 24 h of onset and associated with patient worsening. Clinically it is relevant to detect the patients that will develop HE from their initial computed tomography (CT) scans which could improve patient management and treatment decisions. However, this is a significant challenge due to the predictive nature of the task and its low prevalence, which hinders the availability of large datasets with the required longitudinal information. In this work, we present an end-to-end deep learning framework capable of predicting which cases will exhibit HE using only the initial basal image. We introduce a deep learning framework based on the 2D EfficientNet B0 model to predict the occurrence of HE using initial non-contrasted CT scans and their corresponding lesion annotation as priors. We used an in-house acquired dataset of 122 ICH patients, including 35 HE cases, containing longitudinal CT scans with manual lesion annotations in both basal and follow-up (obtained within 24 h after the basal scan). Experiments were conducted using a 5-fold cross-validation strategy. We addressed the limited data problem by incorporating synthetic images into the training process. To the best of our knowledge, our approach is novel in the field of HE prediction, being the first to use image synthesis to enhance results. We studied different scenarios such as training only with the original scans, using standard image augmentation techniques, and using synthetic image generation. The best performance was achieved by adding five generated versions of each image, along with standard data augmentation, during the training process. This significantly improved ($p=0.0003$) the performance obtained with our baseline model using directly the original CT scans from an Accuracy of 0.56 to 0.84, F1-Score of 0.53 to 0.82, Sensitivity of 0.51 to 0.77, and Specificity of 0.60 to 0.91, respectively. The proposed approach shows promising results in predicting HE, especially with the inclusion of synthetically generated images. The obtained results highlight the significance of this research direction, which has the potential to improve the clinical management of patients with hemorrhagic stroke. The code is available at: https://github.com/NIC-VICOROB/HE-prediction-SynthCT.

CycleGAN has been leveraged to synthesize a CT image from an available MR image after trained on unpaired data. Due to the lack of direct constraints between the synthetic and the input images, CycleGAN cannot guarantee structural consistency and often generates inaccurate mappings that shift the anatomy, which is highly undesirable for downstream clinical applications such as MRI-guided radiotherapy treatment planning and PET/MRI attenuation correction. In this paper, we propose a cycle-consistent and semantics-preserving generative adversarial network, referred as CycleSGAN, for unpaired MR-to-CT image synthesis. Our design features a novel and generic way to incorporate semantic information into CycleGAN. This is done by designing a pair of three-player games within the CycleGAN framework where each three-player game consists of one generator and two discriminators to formulate two distinct types of adversarial learning: appearance adversarial learning and structure adversarial learning. These two types of adversarial learning are alternately trained to ensure both realistic image synthesis and semantic structure preservation. Results on unpaired hip MR-to-CT image synthesis show that our method produces better synthetic CT images in both accuracy and visual quality as compared to other state-of-the-art (SOTA) unpaired MR-to-CT image synthesis methods.

Lung cancer has the highest mortality rate among cancers. The commonly used clinical method for diagnosing lung cancer is the CT-guided percutaneous transthoracic lung biopsy (CT-PTLB), but this method requires a high level of clinical experience from doctors. In this work, an automatic path planning method for CT-PTLB is proposed to provide doctors with auxiliary advice on puncture paths. The proposed method comprises three steps: preprocessing, initial path selection, and path evaluation. During preprocessing, the chest organs required for subsequent path planning are segmented. During the initial path selection, a target point selection method for selecting biopsy samples according to biopsy sampling requirements is proposed, which includes a down-sampling algorithm suitable for different nodule shapes. Entry points are selected according to the selected target points and clinical constraints. During the path evaluation, the clinical needs of lung biopsy surgery are first quantified as path evaluation indicators and then divided according to their evaluation perspective into risk and execution indicators. Then, considering the impact of the correlation between indicators, a path scoring system based on the double spherical constraint Pareto and the importance-correlation degree of the indicators is proposed to evaluate the comprehensive performance of the planned paths. The proposed method is retrospectively tested on 6 CT images and prospectively tested on 25 CT images. The experimental results indicate that the method proposed in this work can be used to plan feasible puncture paths for different cases and can serve as an auxiliary tool for lung biopsy surgery.

This study presents an innovative hybrid deep learning (DL) framework that reformulates the sagittal MRI-based anterior cruciate ligament (ACL) tear classification task as a novelty detection problem to tackle class imbalance. We introduce a highly discriminative novelty score, which leverages the aleatoric semantic uncertainty as this is modeled in the class scores outputted by the YOLOv5-nano object detection (OD) model. To account for tissue continuity, we propose using the global scores (probability vector) when the model is applied to the entire sagittal sequence. The second module of the proposed pipeline constitutes the MINIROCKET timeseries classification model for determining whether a knee has an ACL tear. To better evaluate the generalization capabilities of our approach, we also carry out cross-database testing involving two public databases (KneeMRI and MRNet) and a validation-only database from University General Hospital of Larissa, Greece. Our method consistently outperformed (p-value<0.05) the state-of-the-art (SOTA) approaches on the KneeMRI dataset and achieved better accuracy and sensitivity on the MRNet dataset. It also generalized remarkably good, especially when the model had been trained on KneeMRI. The presented framework generated at least 2.1 times less carbon emissions and consumed at least 2.6 times less energy, when compared with SOTA. The integration of aleatoric semantic uncertainty-based scores into a novelty detection framework, when combined with the use of lightweight OD and timeseries classification models, have the potential to revolutionize the MRI-based injury detection by setting a new precedent in diagnostic precision, speed and environmental sustainability. Our resource-efficient framework offers potential for widespread application.

Photoacoustic tomography (PAT) is a powerful imaging modality for visualizing tissue physiology and exogenous contrast agents. However, PAT faces challenges in visualizing deep-seated vascular structures due to light scattering, absorption, and reduced signal intensity with depth. Optical coherence tomography angiography (OCTA) offers high-contrast visualization of vasculature networks, yet its imaging depth is limited to a millimeter scale. Herein, we propose OCPA-Net, a novel unsupervised deep learning method that utilizes the rich vascular feature of OCTA to enhance PAT images. Trained on unpaired OCTA and PAT images, OCPA-Net incorporates a vessel-aware attention module to enhance deep-seated vessel details captured from OCTA. It leverages a domain-adversarial loss function to enforce structural consistency and a novel identity invariant loss to mitigate excessive image content generation. We validate the structural fidelity of OCPA-Net on simulation experiments, and then demonstrate its vascular enhancement performance on in vivo imaging experiments of tumor-bearing mice and contrast-enhanced pregnant mice. The results show the promise of our method for comprehensive vessel-related image analysis in preclinical research applications.

Automated cervical cancer screening through computer-assisted diagnosis has shown considerable potential to improve screening accessibility and reduce associated costs and errors. However, classification performance on whole slide images (WSIs) remains suboptimal due to patient-specific variations. To improve the precision of the screening, pathologists not only analyze the characteristics of suspected abnormal cells, but also compare them with normal cells. Motivated by this practice, we propose a novel cervical cell comparative learning method that leverages pathologist knowledge to learn the differences between normal and suspected abnormal cells within the same WSI. Our method employs two pre-trained YOLOX models to detect suspected abnormal and normal cells in a given WSI. A self-supervised model then extracts features for the detected cells. Subsequently, a tailored Transformer encoder fuses the cell features to obtain WSI instance embeddings. Finally, attention-based multi-instance learning is applied to achieve classification. The experimental results show an AUC of 0.9319 for our proposed method. Moreover, the method achieved professional pathologist-level performance, indicating its potential for clinical applications.

Reliability learning and interpretable decision-making are crucial for multi-modality medical image segmentation. Although many works have attempted multi-modality medical image segmentation, they rarely explore how much reliability is provided by each modality for segmentation. Moreover, the existing approach of decision-making such as the $softmax$ function lacks the interpretability for multi-modality fusion. In this study, we proposed a novel approach named contextual discounted evidential network (CDE-Net) for reliability learning and interpretable decision-making under multi-modality medical image segmentation. Specifically, the CDE-Net first models the semantic evidence by uncertainty measurement using the proposed evidential decision-making module. Then, it leverages the contextual discounted fusion layer to learn the reliability provided by each modality. Finally, a multi-level loss function is deployed for the optimization of evidence modeling and reliability learning. Moreover, this study elaborates on the framework interpretability by discussing the consistency between pixel attribution maps and the learned reliability coefficients. Extensive experiments are conducted on both multi-modality brain and liver datasets. The CDE-Net gains high performance with an average Dice score of 0.914 for brain tumor segmentation and 0.913 for liver tumor segmentation, which proves CDE-Net has great potential to facilitate the interpretation of artificial intelligence-based multi-modality medical image fusion.

Intracranial aneurysm (IA) is a prevalent disease that poses a significant threat to human health. The use of computed tomography angiography (CTA) as a diagnostic tool for IAs remains time-consuming and challenging. Deep neural networks (DNNs) have made significant advancements in the field of medical image segmentation. Nevertheless, training large-scale DNNs demands substantial quantities of high-quality labeled data, making the annotation of numerous brain CTA scans a challenging endeavor. To address these challenges and effectively develop a robust IAs segmentation model from a large amount of unlabeled training data, we propose a triple learning framework (TLF). The framework primarily consists of three learning paradigms: pseudo-supervised learning, contrastive learning, and confident learning. This paper introduces an enhanced mean teacher model and voxel-selective strategy to conduct pseudo-supervised learning on unreliable labeled training data. Concurrently, we construct the positive and negative training pairs within the high-level semantic feature space to improve the overall learning efficiency of the TLF through contrastive learning. In addition, a multi-scale confident learning is proposed to correct unreliable labels, which enables the acquisition of broader local structural information instead of relying on individual voxels. To evaluate the effectiveness of our method, we conducted extensive experiments on a self-built database of hundreds of cases of brain CTA scans with IAs. Experimental results demonstrate that our method can effectively learn a robust CTA scan-based IAs segmentation model using unreliable labeled data, outperforming state-of-the-art methods in terms of segmentation accuracy. Codes are released at https://github.com/XueShuangqian/TLF.