Visual Computing for Industry Biomedicine and Art最新文献

This study presents a novel visualization approach to explainable artificial intelligence for graph-based visual question answering (VQA) systems. The method focuses on identifying false answer predictions by the model and offers users the opportunity to directly correct mistakes in the input space, thus facilitating dataset curation. The decision-making process of the model is demonstrated by highlighting certain internal states of a graph neural network (GNN). The proposed system is built on top of a GraphVQA framework that implements various GNN-based models for VQA trained on the GQA dataset. The authors evaluated their tool through the demonstration of identified use cases, quantitative measures, and a user study conducted with experts from machine learning, visualization, and natural language processing domains. The authors' findings highlight the prominence of their implemented features in supporting the users with incorrect prediction identification and identifying the underlying issues. Additionally, their approach is easily extendable to similar models aiming at graph-based question answering.

Breast cancer is one of the most common malignancies among women globally. Magnetic resonance imaging (MRI), as the final non-invasive diagnostic tool before biopsy, provides detailed free-text reports that support clinical decision-making. Therefore, the effective utilization of the information in MRI reports to make reliable decisions is crucial for patient care. This study proposes a novel method for BI-RADS classification using breast MRI reports. Large language models are employed to transform free-text reports into structured reports. Specifically, missing category information (MCI) that is absent in the free-text reports is supplemented by assigning default values to the missing categories in the structured reports. To ensure data privacy, a locally deployed Qwen-Chat model is employed. Furthermore, to enhance the domain-specific adaptability, a knowledge-driven prompt is designed. The Qwen-7B-Chat model is fine-tuned specifically for structuring breast MRI reports. To prevent information loss and enable comprehensive learning of all report details, a fusion strategy is introduced, combining free-text and structured reports to train the classification model. Experimental results show that the proposed BI-RADS classification method outperforms existing report classification methods across multiple evaluation metrics. Furthermore, an external test set from a different hospital is used to validate the robustness of the proposed approach. The proposed structured method surpasses GPT-4o in terms of performance. Ablation experiments confirm that the knowledge-driven prompt, MCI, and the fusion strategy are crucial to the model's performance.

Clamping and removal of the nucleus pulposus (NP) are critical operations during transforaminal endoscopic lumbar discectomy. To meet the challenge of simulating the NP in real-time for better training output, an improved material point method is proposed to represent the physical properties of the NP and compute its deformation in real time. Corresponding volume rendering of the NP and its hosting bones are also presented. The virtual operation procedures are then implemented into a training prototype and subsequently tested through simulation experiments and subjective evaluation. The results have demonstrated the feasibility of the approach.

Federated learning (FL) has shown great potential in addressing data privacy issues in medical image analysis. However, varying data distributions across different sites can create challenges in aggregating client models and achieving good global model performance. In this study, we propose a novel personalized contrastive representation FL framework, named PCRFed, which leverages contrastive representation learning to address the non-independent and identically distributed (non-IID) challenge and dynamically adjusts the distance between local clients and the global model to improve each client's performance without incurring additional communication costs. The proposed weighted model-contrastive loss provides additional regularization for local models, optimizing their respective distributions while effectively utilizing information from all clients to mitigate performance challenges caused by insufficient local data. The PCRFed approach was evaluated on two non-IID medical image segmentation datasets, and the results show that it outperforms several state-of-the-art FL frameworks, achieving higher single-client performance while ensuring privacy preservation and minimal communication costs. Our PCRFed framework can be adapted to various encoder-decoder segmentation network architectures and holds significant potential for advancing the use of FL in real-world medical applications. Based on a multi-center dataset, our framework demonstrates superior overall performance and higher single-client performance, achieving a 2.63% increase in the average Dice score for prostate segmentation.

Breast cancer, which is the most commonly diagnosed cancers among women, is a notable health issues globally. Breast cancer is a result of abnormal cells in the breast tissue growing out of control. Histopathology, which refers to the detection and learning of tissue diseases, has appeared as a solution for breast cancer treatment as it plays a vital role in its diagnosis and classification. Thus, considerable research on histopathology in medical and computer science has been conducted to develop an effective method for breast cancer treatment. In this study, a vision Transformer (ViT) was employed to classify tumors into two classes, benign and malignant, in the Breast Cancer Histopathological Database (BreakHis). To enhance the model performance, we introduced the novel multi-head locality large kernel self-attention during fine-tuning, achieving an accuracy of 95.94% at 100× magnification, thereby improving the accuracy by 3.34% compared to a standard ViT (which uses multi-head self-attention). In addition, the application of principal component analysis for dimensionality reduction led to an accuracy improvement of 3.34%, highlighting its role in mitigating overfitting and reducing the computational complexity. In the final phase, SHapley Additive exPlanations, Local Interpretable Model-agnostic Explanations, and Gradient-weighted Class Activation Mapping were used for the interpretability and explainability of machine-learning models, aiding in understanding the feature importance and local explanations, and visualizing the model attention. In another experiment, ensemble learning with VGGIN further boosted the performance to 97.13% accuracy. Our approach exhibited a 0.98% to 17.13% improvement in accuracy compared with state-of-the-art methods, establishing a new benchmark for breast cancer histopathological image classification.

Die-forging structural parts are widely used in the main load-bearing components of aircrafts because of their excellent mechanical properties and fatigue resistance. However, the forming and heat treatment processes of die-forging structural parts are complex, leading to high levels of internal stress and a complex distribution of residual stress fields (RSFs), which affect the deformation, fatigue life, and failure of structural parts throughout their lifecycles. Hence, the global RSF can provide the basis for process control. The existing RSF inference method based on deformation force data can utilize monitoring data to infer the global RSF of a regular part. However, owing to the irregular geometry of die-forging structural parts and the complexity of the RSF, it is challenging to solve ill-conditioned problems during the inference process, which makes it difficult to obtain the RSF accurately. This paper presents a global RSF inference method for the die-forging structural parts based on the fusion of monitoring data and distribution prior. Prior knowledge was derived from the RSF distribution trends obtained through finite element analysis. This enables the low-dimensional characterization of the RSF, reducing the number of parameters required to solve the equations. The effectiveness of this method was validated in both simulation and actual environments.

Cataract is the leading ocular disease of blindness and visual impairment globally. Deep neural networks (DNNs) have achieved promising cataracts recognition performance based on anterior segment optical coherence tomography (AS-OCT) images; however, they have poor explanations, limiting their clinical applications. In contrast, visual features extracted from original AS-OCT images and their transform forms (e.g., AS-OCT-based histograms) have good explanations but have not been fully exploited. Motivated by these observations, an explainable machine learning framework to recognize cataracts severity levels automatically using AS-OCT images was proposed, consisting of three stages: visual feature extraction, feature importance explanation and selection, and recognition. First, the intensity histogram and intensity-based statistical methods are applied to extract visual features from original AS-OCT images and AS-OCT-based histograms. Subsequently, the SHapley Additive exPlanations and Pearson correlation coefficient methods are applied to analyze the feature importance and select significant visual features. Finally, an ensemble multi-class ridge regression method is applied to recognize the cataracts severity levels based on the selected visual features. Experiments on a clinical AS-OCT-NC dataset demonstrate that the proposed framework not only achieves competitive performance through comparisons with DNNs, but also has a good explanation ability, meeting the requirements of clinical diagnostic practice.

Fluorescence endoscopy technology utilizes a light source of a specific wavelength to excite the fluorescence signals of biological tissues. This capability is extremely valuable for the early detection and precise diagnosis of pathological changes. Identifying a suitable experimental approach and metric for objectively and quantitatively assessing the imaging quality of fluorescence endoscopy is imperative to enhance the image evaluation criteria of fluorescence imaging technology. In this study, we propose a new set of standards for fluorescence endoscopy technology to evaluate the optical performance and image quality of fluorescence imaging objectively and quantitatively. This comprehensive set of standards encompasses fluorescence test models and imaging quality assessment protocols to ensure that the performance of fluorescence endoscopy systems meets the required standards. In addition, it aims to enhance the accuracy and uniformity of the results by standardizing testing procedures. The formulation of pivotal metrics and testing methodologies is anticipated to facilitate direct quantitative comparisons of the performance of fluorescence endoscopy devices. This advancement is expected to foster the harmonization of clinical and preclinical evaluations using fluorescence endoscopy imaging systems, thereby improving diagnostic precision and efficiency.

The vision transformer (ViT) architecture, with its attention mechanism based on multi-head attention layers, has been widely adopted in various computer-aided diagnosis tasks due to its effectiveness in processing medical image information. ViTs are notably recognized for their complex architecture, which requires high-performance GPUs or CPUs for efficient model training and deployment in real-world medical diagnostic devices. This renders them more intricate than convolutional neural networks (CNNs). This difficulty is also challenging in the context of histopathology image analysis, where the images are both limited and complex. In response to these challenges, this study proposes a TokenMixer hybrid-architecture that combines the strengths of CNNs and ViTs. This hybrid architecture aims to enhance feature extraction and classification accuracy with shorter training time and fewer parameters by minimizing the number of input patches employed during training, while incorporating tokenization of input patches using convolutional layers and encoder transformer layers to process patches across all network layers for fast and accurate breast cancer tumor subtype classification. The TokenMixer mechanism is inspired by the ConvMixer and TokenLearner models. First, the ConvMixer model dynamically generates spatial attention maps using convolutional layers, enabling the extraction of patches from input images to minimize the number of input patches used in training. Second, the TokenLearner model extracts relevant regions from the selected input patches, tokenizes them to improve feature extraction, and trains all tokenized patches in an encoder transformer network. We evaluated the TokenMixer model on the BreakHis public dataset, comparing it with ViT-based and other state-of-the-art methods. Our approach achieved impressive results for both binary and multi-classification of breast cancer subtypes across various magnification levels (40×, 100×, 200×, 400×). The model demonstrated accuracies of 97.02% for binary classification and 93.29% for multi-classification, with decision times of 391.71 and 1173.56 s, respectively. These results highlight the potential of our hybrid deep ViT-CNN architecture for advancing tumor classification in histopathological images. The source code is accessible: https://github.com/abimouloud/TokenMixer .

With the exponential rise in global air traffic, ensuring swift passenger processing while countering potential security threats has become a paramount concern for aviation security. Although X-ray baggage monitoring is now standard, manual screening has several limitations, including the propensity for errors, and raises concerns about passenger privacy. To address these drawbacks, researchers have leveraged recent advances in deep learning to design threat-segmentation frameworks. However, these models require extensive training data and labour-intensive dense pixel-wise annotations and are finetuned separately for each dataset to account for inter-dataset discrepancies. Hence, this study proposes a semi-supervised contour-driven broad learning system (BLS) for X-ray baggage security threat instance segmentation referred to as C-BLX. The research methodology involved enhancing representation learning and achieving faster training capability to tackle severe occlusion and class imbalance using a single training routine with limited baggage scans. The proposed framework was trained with minimal supervision using resource-efficient image-level labels to localize illegal items in multi-vendor baggage scans. More specifically, the framework generated candidate region segments from the input X-ray scans based on local intensity transition cues, effectively identifying concealed prohibited items without entire baggage scans. The multi-convolutional BLS exploits the rich complementary features extracted from these region segments to predict object categories, including threat and benign classes. The contours corresponding to the region segments predicted as threats were then utilized to yield the segmentation results. The proposed C-BLX system was thoroughly evaluated on three highly imbalanced public datasets and surpassed other competitive approaches in baggage-threat segmentation, yielding 90.04%, 78.92%, and 59.44% in terms of mIoU on GDXray, SIXray, and Compass-XP, respectively. Furthermore, the limitations of the proposed system in extracting precise region segments in intricate noisy settings and potential strategies for overcoming them through post-processing techniques were explored (source code will be available at https://github.com/Divs1159/CNN_BLS .).