Manual Annotation Research Articles

Automated detection of Bornean white-bearded gibbon (Hylobates albibarbis) vocalizations using an open-source framework for deep learning.

Passive acoustic monitoring is a promising tool for monitoring at-risk populations of vocal species, yet, extracting relevant information from large acoustic datasets can be time-consuming, creating a bottleneck at the point of analysis. To address this, an open-source framework for deep learning in bioacoustics to automatically detect Bornean white-bearded gibbon (Hylobates albibarbis) "great call" vocalizations in a long-term acoustic dataset from a rainforest location in Borneo is adapted. The steps involved in developing this solution are described, including collecting audio recordings, developing training and testing datasets, training neural network models, and evaluating model performance. The best model performed at a satisfactory level (F score = 0.87), identifying 98% of the highest-quality calls from 90 h of manually annotated audio recordings and greatly reduced analysis times when compared to a human observer. No significant difference was found in the temporal distribution of great call detections between the manual annotations and the model's output. Future work should seek to apply this model to long-term acoustic datasets to understand spatiotemporal variations in H. albibarbis' calling activity. Overall, a roadmap is presented for applying deep learning to identify the vocalizations of species of interest, which can be adapted for monitoring other endangered vocalizing species.

Deep Location Soft-Embedding-Based Network With Regional Scoring for Mammogram Classification.

Early detection and treatment of breast cancer can significantly reduce patient mortality, and mammogram is an effective method for early screening. Computer-aided diagnosis (CAD) of mammography based on deep learning can assist radiologists in making more objective and accurate judgments. However, existing methods often depend on datasets with manual segmentation annotations. In addition, due to the large image sizes and small lesion proportions, many methods that do not use region of interest (ROI) mostly rely on multi-scale and multi-feature fusion models. These shortcomings increase the labor, money, and computational overhead of applying the model. Therefore, a deep location soft-embedding-based network with regional scoring (DLSEN-RS) is proposed. DLSEN-RS is an end-to-end mammography image classification method containing only one feature extractor and relies on positional embedding (PE) and aggregation pooling (AP) modules to locate lesion areas without bounding boxes, transfer learning, or multi-stage training. In particular, the introduced PE and AP modules exhibit versatility across various CNN models and improve the model's tumor localization and diagnostic accuracy for mammography images. Experiments are conducted on published INbreast and CBIS-DDSM datasets, and compared to previous state-of-the-art mammographic image classification methods, DLSEN-RS performed satisfactorily.

3D Vessel Segmentation With Limited Guidance of 2D Structure-Agnostic Vessel Annotations.

Delineating 3D blood vessels of various anatomical structures is essential for clinical diagnosis and treatment, however, is challenging due to complex structure variations and varied imaging conditions. Although recent supervised deep learning models have demonstrated their superior capacity in automatic 3D vessel segmentation, the reliance on expensive 3D manual annotations and limited capacity for annotation reuse among different vascular structures hinder their clinical applications. To avoid the repetitive and costly annotating process for each vascular structure and make full use of existing annotations, this paper proposes a novel 3D shape-guided local discrimination (3D-SLD) model for 3D vascular segmentation under limited guidance from public 2D vessel annotations. The primary hypothesis is that 3D vessels are composed of semantically similar voxels and often exhibit tree-shaped morphology. Accordingly, the 3D region discrimination loss is firstly proposed to learn the discriminative representation measuring voxel-wise similarities and cluster semantically consistent voxels to form the candidate 3D vascular segmentation in unlabeled images. Secondly, the shape distribution from existing 2D structure-agnostic vessel annotations is introduced to guide the 3D vessels with the tree-shaped morphology by the adversarial shape constraint loss. Thirdly, to enhance training stability and prediction credibility, the highlighting-reviewing-summarizing (HRS) mechanism is proposed. This mechanism involves summarizing historical models to maintain temporal consistency and identifying credible pseudo labels as reliable supervision signals. Only guided by public 2D coronary artery annotations, our method achieves results comparable to SOTA barely-supervised methods in 3D cerebrovascular segmentation, and the best DSC in 3D hepatic vessel segmentation, demonstrating the effectiveness of our method.

A novel pipeline employing deep multi-attention channels network for the autonomous detection of metastasizing cells through fluorescence microscopy

Metastasis driven by cancer cell migration is the leading cause of cancer-related deaths. It involves significant changes in the organization of the cytoskeleton, which includes the actin microfilaments and the vimentin intermediate filaments. Understanding how these filament change cells from normal to invasive offers insights that can be used to improve cancer diagnosis and therapy. We have developed a computational, transparent, large-scale and imaging-based pipeline, that can distinguish between normal human cells and their isogenically matched, oncogenically transformed, invasive and metastasizing counterparts, based on the spatial organization of actin and vimentin filaments in the cell cytoplasm. Due to the intricacy of these subcellular structures, manual annotation is not trivial to automate. We used established deep learning methods and our new multi-attention channel architecture. To ensure a high level of interpretability of the network, which is crucial for the application area, we developed an interpretable global explainable approach correlating the weighted geometric mean of the total cell images and their local GradCam scores. The methods offer detailed, objective and measurable understanding of how different components of the cytoskeleton contribute to metastasis, insights that can be used for future development of novel diagnostic tools, such as a nanometer level, vimentin filament-based biomarker for digital pathology, and for new treatments that significantly can increase patient survival.

Quantitative Analysis of Small Intestinal Motility in 3D Cine-MRI Using Centerline-Aware Motion Estimation.

Currently available tools for noninvasive motility quantification of the small intestine are limited to dynamic 2D MRI scans, which are limited in their ability to differentiate between types of intestinal motility. To develop a method for quantification and characterization of small intestinal motility in 3D, capable of differentiating motile, non-motile and peristaltic motion patterns. Prospective. Fourteen healthy volunteers (127 small intestinal segments) and 10 patients with Crohn's disease (87 small intestinal segments). 3.0 T, 3D balanced fast field echo sequence, 1 volume per second. Using deformable image registration between subsequent volumes, the local velocity within the intestinal lumen was quantified. Average velocity and average absolute velocity along intestinal segments were used with linear classifiers to differentiate motile from non-motile intestines, as well as erratic motility from peristalsis. The mean absolute velocity of small intestinal content was compared between healthy volunteers and Crohn's disease patients, and the discriminative power of the proposed motility metrics for detecting motility and peristalsis was determined. The consensus of two observers was used as referenced standard. Student's t-test to assess differences between groups; area under the receiver operating characteristic curve (AUC) to assess discriminative ability. P < 0.001 was considered significant. A significant difference in the absolute velocity of intestinal content between Crohn's patients and healthy volunteers was observed (median [IQR] 1.06 [0.61, 1.56] mm/s vs. 1.84 [1.37, 2.43] mm/s), which was consistent with manual reference annotations of motile activity. The proposed method had a strong discriminative performance for detecting non-motile intestines (AUC 0.97) and discernible peristalsis (AUC 0.81). Analysis of 3D cine-MRI using centerline-aware motion estimation has the potential to allow noninvasive characterization of small intestinal motility and peristaltic motion in 3D. 3 TECHNICAL EFFICACY: Stage 2.

High-resolution segmentations of the hypothalamus and its subregions for training of segmentation models

Segmentation of brain structures on magnetic resonance imaging (MRI) is a highly relevant neuroimaging topic, as it is a prerequisite for different analyses such as volumetry or shape analysis. Automated segmentation facilitates the study of brain structures in larger cohorts when compared with manual segmentation, which is time-consuming. However, the development of most automated methods relies on large and manually annotated datasets, which limits the generalizability of these methods. Recently, new techniques using synthetic images have emerged, reducing the need for manual annotation. Here we provide a dataset composed of label maps built from publicly available ultra-high resolution ex vivo MRI from 10 whole hemispheres, which can be used to develop segmentation methods using synthetic data. The label maps are obtained with a combination of manual labels for the hypothalamic regions and automated segmentations for the rest of the brain, and mirrored to simulate entire brains. We also provide the pre-processed ex vivo scans, as this dataset can support future projects to include other structures after these are manually segmented.

SPORT: A Subgraph Perspective on Graph Classification with Label Noise

Graph neural networks (GNNs) have achieved great success recently on graph classification tasks using supervised end-to-end training. Unfortunately, extensive noisy graph labels could exist in the real world because of the complicated processes of manual graph data annotations, which may significantly degrade the performance of GNNs. Therefore, we investigate the problem of graph classification with label noise, which is demanding because of the complex graph representation learning issue and serious memorization of noisy samples. In this work, we present a novel approach called S ubgra p h Set Netw or k with Sample Selection and Consis t ency Learning (SPORT) for this problem. To release the overfitting of GNNs, SPORT proposes to characterize each graph as a set of subgraphs generated by certain predefined stratagems, which can be viewed as samples from its underlying semantic distribution in graph space. Then we develop an equivariant network to encode the subgraph set with the consideration of the symmetry group. To further release the influences of noisy examples, we leverage the predictions of subgraphs to measure the likelihood of a sample being clean or noisy, followed by effective label updating. In addition, we propose a joint loss to advance the model generalizability by introducing consistency regularization. Comprehensive experiments on a wide range of graph classification datasets demonstrate the effectiveness of our SPORT. Specifically, SPORT outperforms the most competing baseline by up to 6.4%.

Pretrained subtraction and segmentation model for coronary angiograms

This study introduces a novel self-supervised learning method for single-frame subtraction and vessel segmentation in coronary angiography, addressing the scarcity of annotated medical samples in AI applications. We pretrain a U-Net model on a large dataset of unannotated coronary angiograms using an image-to-image translation framework, then fine-tune it on a limited set of manually annotated samples. The pretrained model excels at comprehensive single-frame subtraction, outperforming existing DSA methods. Fine-tuning with just 40 samples yields a Dice coefficient of 0.828 for vessel segmentation. On the public XCAD dataset, our model sets a new state-of-the-art benchmark with a Dice coefficient of 0.755, surpassing both unsupervised and supervised learning approaches. This method achieves robust single-frame subtraction and demonstrates that combining pretraining with minimal fine-tuning enables accurate coronary vessel segmentation with limited manual annotations. We successfully apply this approach to assist physicians in visualizing potential vascular stenosis sites during coronary angiography. Code, dataset, and a live demo will be available available at: https://github.com/newfyu/DeepSA.

The Fine-Tuned Large Language Model for Extracting the Progressive Bone Metastasis from Unstructured Radiology Reports.

Early detection of patients with impending bone metastasis is crucial for prognosis improvement. This study aimed to investigate the feasibility of a fine-tuned, locally run large language model (LLM) in extracting patients with bone metastasis in unstructured Japanese radiology report and to compare its performance with manual annotation. This retrospective study included patients with "metastasis" in radiological reports (April 2018-January 2019, August-May 2022, and April-December 2023 for training, validation, and test datasets of 9559, 1498, and 7399 patients, respectively). Radiologists reviewed the clinical indication and diagnosis sections of the radiological report (used as input data) and classified them into groups 0 (no bone metastasis), 1 (progressive bone metastasis), and 2 (stable or decreased bone metastasis). The data for group 0 was under-sampled in training and test datasets due to group imbalance. The best-performing model from the validation set was subsequently tested using the testing dataset. Two additional radiologists (readers 1 and 2) were involved in classifying radiological reports within the test dataset for testing purposes. The fine-tuned LLM, reader 1, and reader 2 demonstrated an accuracy of 0.979, 0.996, and 0.993, sensitivity for groups 0/1/2 of 0.988/0.947/0.943, 1.000/1.000/0.966, and 1.000/0.982/0.954, and time required for classification (s) of 105, 2312, and 3094 in under-sampled test dataset (n = 711), respectively. Fine-tuned LLM extracted patients with bone metastasis, demonstrating satisfactory performance that was comparable to or slightly lower than manual annotation by radiologists in a noticeably shorter time.

Learning from the few: Fine-grained approach to pediatric wrist pathology recognition on a limited dataset

Wrist pathologies, particularly fractures common among children and adolescents, present a critical diagnostic challenge. While X-ray imaging remains a prevalent diagnostic tool, the increasing misinterpretation rates highlight the need for more accurate analysis, especially considering the lack of specialized training among many surgeons and physicians. Recent advancements in deep convolutional neural networks offer promise in automating pathology detection in trauma X-rays. However, distinguishing subtle variations between pediatric wrist pathologies in X-rays remains challenging. Traditional manual annotation, though effective, is laborious, costly, and requires specialized expertise. In this paper, we address the challenge of pediatric wrist pathology recognition with a fine-grained approach, aimed at automatically identifying discriminative regions in X-rays without manual intervention. We refine our fine-grained architecture through ablation analysis and the integration of LION. Leveraging Grad-CAM, an explainable AI technique, we highlight these regions. Despite using limited data, reflective of real-world medical study constraints, our method consistently outperforms state-of-the-art image recognition models on both augmented and original (challenging) test sets. Our proposed refined architecture achieves an increase in accuracy of 1.06% and 1.25% compared to the baseline method, resulting in accuracies of 86% and 84%, respectively. Moreover, our approach demonstrates the highest fracture sensitivity of 97%, highlighting its potential to enhance wrist pathology recognition.

Mapping the Multiscale Proteomic Organization of Cellular and Disease Phenotypes.

While the primary sequences of human proteins have been cataloged for over a decade, determining how these are organized into a dynamic collection of multiprotein assemblies, with structures and functions spanning biological scales, is an ongoing venture. Systematic and data-driven analyses of these higher-order structures are emerging, facilitating the discovery and understanding of cellular phenotypes. At present, knowledge of protein localization and function has been primarily derived from manual annotation and curation in resources such as the Gene Ontology, which are biased toward richly annotated genes in the literature. Here, we envision a future powered by data-driven mapping of protein assemblies. These maps can capture and decode cellular functions through the integration of protein expression, localization, and interaction data across length scales and timescales. In this review, we focus on progress toward constructing integrated cell maps that accelerate the life sciences and translational research.

Generating realistic training images from synthetic data for excavator pose estimation

Computer vision-based 3D pose estimation for automated excavator operation monitoring requires numerous training images annotated with 3D pose labels. Owing to challenges in collecting such datasets in a field setting, using synthetic images from virtual environments has emerged recently. However, synthetic images lack the realism inherent in onsite images, potentially impacting pose estimation performance on real images. This paper thus proposes a generative model for generating realistic training excavator images with multiple backgrounds. The evaluation was conducted by comparing estimation models trained on synthetic images (Model #1), generated excavator images with single background (Model #2), and generated excavator images with multiple backgrounds (Model #3). Model #3 exhibited the lowest mean angular error of 5.96° on real data, implying its superiority in generalizing real patterns. The proposed model facilitates data acquisition for improving pose estimation without manual annotation, providing rich information on excavator movements for proactive safety and productivity management.

Classifiers of Data Sharing Statements in Clinical Trial Records.

Digital individual participant data (IPD) from clinical trials are increasingly distributed for potential scientific reuse. The identification of available IPD, however, requires interpretations of textual data-sharing statements (DSS) in large databases. Recent advancements in computational linguistics include pre-trained language models that promise to simplify the implementation of effective classifiers based on textual inputs. In a subset of 5,000 textual DSS from ClinicalTrials.gov, we evaluate how well classifiers based on domain-specific pre-trained language models reproduce original availability categories as well as manually annotated labels. Typical metrics indicate that classifiers that predicted manual annotations outperformed those that learned to output the original availability categories. This suggests that the textual DSS descriptions contain applicable information that the availability categories do not, and that such classifiers could thus aid the automatic identification of available IPD in large trial databases.

Automatic Classification of Conclusions from Multi-Tracer Reports of PET Brain Imaging in Cognitive Impairment.

The goal of this paper is to build an automatic way to interpret conclusions from brain molecular imaging reports performed for investigation of cognitive disturbances (FDG, Amyloid and Tau PET) by comparing several traditional machine learning (ML) techniques-based text classification methods. Two purposes are defined: to identify positive or negative results in all three modalities, and to extract diagnostic impressions for Alzheimer's Disease (AD), Fronto-Temporal Dementia (FTD), Lewy Bodies Dementia (LBD) based on metabolism of perfusion patterns. A dataset was created by manual parallel annotation of 1668 conclusions of reports from the Nuclear Medicine and Molecular Imaging Division of Geneva University Hospitals. The 6 Machine Learning (ML) algorithms (Support Vector Machine (Linear and Radial Basis function), Naive Bayes, Logistic Regression, Random Forrest, and K-Nearest Neighbors) were trained and evaluated with a 5-fold cross-validation scheme to assess their performance and generalizability. The best classifier was SVM showing the following accuracies: FDG (0.97), Tau (0.94), Amyloid (0.98), Oriented Diagnostic (0.87 for a diagnosis among AD, FTD, LBD, undetermined, other), paving the way for a paradigm shift in the field of data handling in nuclear medicine research.

Understanding and mitigating dimensional collapse of Graph Contrastive Learning: A non-maximum removal approach

Graph Contrastive Learning (GCL) generates graph-level embeddings by maximizing Mutual Information between different augmented views of the same graph (positive pairs), and shows promising performance in graph representation learning (GRL) without the supervision of manual annotations. However, GCL suffers from a dimensional collapse problem, i.e., embedding vectors reside in a restricted low-dimensional subspace, curtailing the expressiveness of certain embedding dimensions. In this paper, we present a theoretical analysis identifying the smoothing effect of graph pooling and graph convolution’s implicit regularization as principal causes of dimension collapse in graph contrastive learning. To mitigate the above effects, we propose a Non-Maximum Removal Graph Contrastive Learning (nmrGCL) approach, which removes “prominent” dimensions (those significantly contributing to the similarity measure) for positive pairs within the pretext task. Comprehensive experiments on multiple benchmark datasets are conducted and the results show that the proposed nmrGCL outperforms the state-of-the-art methods.

Transformer- and joint learning-based dual-domain networks for undersampled MRI segmentation.

Recently, magnetic resonance imaging (MRI) has become a crucial medical imaging technology widely used in clinical practice. However, MRI faces challenges such as the lengthy acquisition time for k-space data and the need for time-consuming manual annotation by radiologists. Traditionally, these challenges have been addressed individually through undersampled MRI reconstruction and automatic segmentation algorithms. Whether undersampled MRI segmentation can be enhanced by treating undersampled MRI reconstruction and segmentation as an end-to-end task, trained simultaneously, rather than as serial tasks should beexplored. We introduce a novel Transformer- and Joint Learning-based Dual-domain Network (TJLD-Net) for undersampled MRIsegmentation. This method significantly enhances feature recognition in the segmentation process by fully utilizing the rich detail obtained during the image reconstruction phase. Consequently, the method can achieve precise and reliable image segmentation even with undersampled k-space data. Additionally, it incorporates an attention mechanism for feature enhancement, which improves the representation of shared features by learning the contextual information in MRimages. Simulation experiments demonstrate that the segmentation performance of TJLD-Net on three datasets is significantly higher than that of the joint model (RecSeg) and six baseline models (where reconstruction and segmentation are regarded as serial tasks). On the CHAOS dataset, the Dice scores of TJLD-Net are, on average, 9.87%, 2.17%, 1.90%, 1.80%, 9.60%, 0.80%, and 6.50% higher than those of the seven compared models. On the ATLAS challenge dataset, the average Dice scores of TJLD-Net improve by 4.23%, 5.63%, 2.30%, 1.53%, 3.57%, 0.93%, and 6.60%. Similarly, on the SKM-TEA dataset, the average Dice scores of TJLD-Net improve by 4.73%, 12.80%, 14.83%, 8.67%, 4.53%, 11.60%, and 12.10%. The novel TJLD-Net model provides a promising solution for undersampled MRI segmentation, overcoming the poor performance issues encountered by automated segmentation algorithms in low-quality acceleratedimaging.

DART: An automated end-to-end object detection pipeline with data Diversification, open-vocabulary bounding box Annotation, pseudo-label Review, and model Training

Accurate real-time object detection is vital across numerous industrial applications, from safety monitoring to quality control. Traditional approaches, however, are hindered by arduous manual annotation and data collection, struggling to adapt to ever-changing environments and novel target objects. To address these limitations, this paper presents DART, an innovative automated end-to-end pipeline that revolutionizes object detection workflows from data collection to model evaluation. It eliminates the need for laborious human labeling and extensive data collection while achieving outstanding accuracy across diverse scenarios. DART encompasses four key stages: (1) Data Diversification using subject-driven image generation (DreamBooth with SDXL), (2) Annotation via open-vocabulary object detection (Grounding DINO) to generate bounding box and class labels, (3) Review of generated images and pseudo-labels by large multimodal models (InternVL-1.5 and GPT-4o) to guarantee credibility, and (4) Training of real-time object detectors (YOLOv8 and YOLOv10) using the verified data. We apply DART to a self-collected dataset of construction machines named Liebherr Product, which contains over 15K high-quality images across 23 categories. The current instantiation of DART significantly increases average precision (AP) from 0.064 to 0.832. Its modular design ensures easy exchangeability and extensibility, allowing for future algorithm upgrades, seamless integration of new object categories, and adaptability to customized environments without manual labeling and additional data collection. The code and dataset are released at https://github.com/chen-xin-94/DART.

LEST: Large language models and spatio-temporal data analysis for enhanced Sino-US exchange rate forecasting

Exchange rate forecasting has a significant impact on a country’s balance of international payments, financial security and stability. Under the context of Sino-US trade war, a series of news events have happened thus increasing the abnormal fluctuations in the RMB exchange rate, and these fluctuations have made it more difficult to forecast the exchange rate. The analysis of news events’ impact on exchange rate fluctuations focused on assigning emotional labels to different economic news and political events by building an emotional lexicon. However, the previous method of using an offline emotional lexicon to analyze events could not be updated promptly making it difficult to better exploit implicit information of news events. Also, the offline emotional lexicon requires manual annotation of events which is unavoidable and subjective and therefore influences the sentiment analysis of the event. Therefore, we build a fusion forecasting framework of spatio-temporal and emotional information for accurate, real-time exchange rate forecasting. First, in the preprocessing phase, we leveraged the situational learning capabilities of Large Language Models (LLMs), specifically ChatGPT, combined with prompt engineering to enable ChatGPT to provide real-time, online, and precise automated sentiment analyses of Sino-US trade events. Second, we proposed the SG-TL fusion model, which integrates Spatio-Temporal Graph Convolutional Networks (STGCN) and Gated Recurrent Unit (GRU) models in the feature processing approach to accurately extract exchange rate features in spatial and temporal dimensions. We also employed Transformer and Long Short-Term Memory (LSTM) models in the forecasting phase to forecast the Sino-US exchange rates. Finally, Experiment results demonstrate that the exchange rate forecasting framework LEST can accurately capture the exchange rate fluctuations during the period of Sino-US trade war. Furthermore, its forecasting performance is more precise, efficient, and stable than existing works. The source code of our developed method is publicly available at https://gitee.com/andyham_andy.ham/lest-forecasting-framework.

A transfer learning-based network model integrating kernel convolution with graph attention mechanism for point cloud segmentation of livestock

Non-contact body size measurement has become a hot research topic in intelligent livestock farming. In regard to body size measurement of large livestock, such as cattle, collecting a substantial number of point clouds is frequently involved. The direct calculation of all point clouds for body size measurement can be impacted as point clouds of different body parts may interfere with each other, which poses huge challenges for the positioning of key points and induces inaccurate positioning, resulting in measurement errors. The accuracy of body size measurement can be improved by segmenting point clouds of different body parts from each other, key measurement points can be precisely located, thus enhancing the accuracy of body size measurement. In this paper, we propose a network model initially trained for pig point cloud segmentation based on the Kernel Convolution integrated with Graph Attention Mechanism (KCGATNet for short), which, through transfer learning techniques, can also be used to achieve successful segmentation of various cattle point clouds using only 7 training samples. The model utilizes two core modules, Kernel Convolution (KC) and Point-based Graph Attention Mechanism (P-GAT), to extract local neighborhood features of point clouds. When using pig body point clouds as training data, it achieved precise segmentation of the head, ears, limbs, torso, and tail of pigs through a downsampling-upsampling architecture. On the test set of pig point clouds, Overall Accuracy (OA) reached 98.1% and mean Intersection over Union (mIoU) was up to 90.5%. Furthermore, when this model served as a pre-trained model and underwent transfer learning using 7 sets of annotated data of Simmental cattle, it achieved a mIoU of 90.1% on a test set of 93 Simmental cattle, 89.6% on a test set of 439 dairy buffalo, 90.2% on a test set of 103 Hereford cattle, and 90.0% on a test set of 119 Black Angus cattle. The experimental outcomes fully demonstrate the robustness of the proposed livestock point cloud segmentation model, KCGATNet. With transfer learning of a small sample size, it can reliably perform point cloud segmentation on a wide range of different breeds of quadrupedal livestock, saving a significant amount of time spent on manual annotation and improving the efficiency of livestock point cloud segmentation models.

Dynamic 4D facial capture pipeline with appearance driven progressive retopology based on optical flow

This paper presents a production-oriented 4D facial reconstruction pipeline designed to produce high-fidelity facial mesh sequences with a consistently structured topology, while preserving the wireframe structure specified by artists. We have designed and developed a compact, efficient, and fast optical capture system based on synchronized camera arrays for high-precision dynamic 3D facial imaging. Unlike prevailing methods that primarily concentrate on single-frame reconstruction, often reliant on labor-intensive manual annotation, our framework exploits the constraint of appearance consistency to autonomously establish feature correspondence and uphold temporal coherence within the mesh. Consequently, our approach eliminates mesh drifting and jitter, enabling full parallelization for dynamic facial expression capture. The proposed pipeline decouples the non-linear deformation of facial expressions from the rigid movements of the skull through a stable external device. Leveraging progressive retopology, our methodology employs artist-guided templates as priors, ensuring the preservation of wireframe structures across the result sequence. Progressive retopology is achieved by constraining different fine-grained features of 3D landmarks, scan surface shapes, and appearance textures. The results of our study showcase facial mesh sequences with production-quality topology, adept at faithfully reproducing character expressions from photographs while achieving artist-friendly stable facial movements.