Annotated Training Research Articles

A novel cross-modal data augmentation method based on contrastive unpaired translation network for kidney segmentation in ultrasound imaging.

Kidney ultrasound (US) image segmentation is one of the key steps in computer-aided diagnosis and treatment planning of kidney diseases. Recently, deep learning (DL) technology has demonstrated promising prospects in automatic kidney US segmentation. However, due to the poor quality, particularly the weak boundaries in kidney US imaging, obtaining accurate annotations for DL-based segmentation methods remain a challenging and time-consuming task. This issue can hinder the application of data-hungry deep learning methods. In this paper, we explore a novel cross-modal data augmentation method aimed at enhancing the performance of DL-based segmentation networks on the limited labeled kidney US dataset. In particular, we adopt a novel method based on contrastive unpaired translation network (CUT) to obtain simulated labeled kidney US images at a low cost from labeled abdomen computed tomography (CT) data and unlabeled kidney US images. To effectively improve the segmentation network performance, we propose an instance-weighting training strategy that simultaneously captures useful information from both the simulated and real labeled kidney US images. We trained our generative networks on a dataset comprising 4418 labeled CT slices and 4594 unlabeled US images. For segmentation network, we used a dataset consisting of 4594 simulated and 100 real kidney US images for training, 20 images for validation, and 169 real images for testing. We compared the performance of our method to several state-of-the-art approaches using the Wilcoxon signed-rank test, and applied the Bonferroni method for multiple comparison correction. The experimental results show that we can synthesize accurate labeled kidney US images with a Fréchet inception distance of 52.52. Moreover, the proposed method achieves a segmentation accuracy of 0.9360±0.0398 for U-Net on normal kidney US images, and 0.7719±0.2449 on the abnormal dataset, as measured by the dice similarity coefficient. When compared to other training strategies, the proposed method demonstrated statistically significant superiority, with all p-values being less than 0.01. The proposed method can effectively improve the accuracy and generalization ability of kidney US image segmentation models with limited annotated training data.

P-1640. Antibiotic prescribing guideline-concordance for sinusitis in adult and pediatric primary care practices

Abstract Background Antibiotics are only sometimes indicated for acute sinusitis, a commonly diagnosed infection in the ambulatory setting and an important target for antibiotic stewardship. However, assessment of the appropriateness of antibiotic use for sinusitis is based on clinical signs and symptoms, and therefore requires manual chart abstraction. We present a study assessing the appropriateness of antibiotic prescribing in adult and pediatric sinusitis visits. Methods Outpatient encounters for sinusitis in adult and pediatric primary care practices within two health systems with an antibiotic prescribed from July 1, 2017 through June 30, 2021 were identified by ICD-10 code (J01 and J32). Of these, 600 encounters were randomly selected for review (300 from adult practices; 300 from pediatric practices). Five trained physician annotators reviewed prescriber notes from these encounters and identified relevant signs/symptoms, duration of illness, and disease trend and severity; annotators used these findings to assess overall guideline concordance of antibiotic prescribing using definitions adapted from clinical guidelines (table 1). Two physicians performed adjudication for disagreements (LD, KWH). Results Demographics and characteristics of included encounters are reported in Table 2. In adults, 19 encounters were excluded after review due to a note being unrelated to the sinusitis encounter; of the remaining 281, 37 (13.2%) were considered guideline concordant, 157 (55.9%) were considered not guideline-concordant, and 87 (31.0%) were potentially guideline concordant but with insufficient information. In pediatrics, 184 (61.3%) were considered guideline concordant, 99 (33.0%) were considered not guideline-concordant, and 17 (5.7%) were potentially guideline concordant but with insufficient information. Common reasons for guideline non-concordance are described in Table 3. Conclusion Antibiotic prescribing forsinusitis was frequently not concordant with guidelines. Prescribing was more frequently concordant in pediatric compared to adult-care settings. In both groups, duration of symptoms was frequently documented as fewer than 10 days, suggesting a potential target for antibiotic stewardship interventions. Disclosures All Authors: No reported disclosures

Cell-APP: A generalizable method for microscopic cell annotation, segmentation, and classification.

High throughput fluorescence microscopy is an essential tool in systems biological studies of eukaryotic cells. Its power can be fully realized when all cells in a field of view and the entire time series can be accurately localized and quantified. These tasks can be mapped to the common paradigm in computer vision: instance segmentation. Recently, supervised deep learning-based methods have become state-of-the-art for cellular instance segmentation. However, these methods require large amounts of high-quality training data. This requirement challenges our ability to train increasingly performant object detectors due to the limited availability of annotated training data, which is typically assembled via laborious hand annotation. Here, we present a generalizable method for generating large instance segmentation training datasets for tissue-culture cells in transmitted light microscopy images. We use datasets created by this method to train vision transformer (ViT) based Mask-RCNNs (Region-based Convolutional Neural Networks) that produce instance segmentations wherein cells are classified as "m-phase" (dividing) or "interphase" (non-dividing). While training these models, we also address the dataset class imbalance between m-phase and interphase cell annotations, which arises for biological reasons, using probabilistically weighted loss functions and partisan training data collection methods. We demonstrate the validity of these approaches by producing highly accurate object detectors that can serve as general tools for the segmentation and classification of morphologically diverse cells. Since the methodology depends only on generic cellular features, we hypothesize that it can be further generalized to most adherent tissue culture cell lines.

Agentic but not Aggressive: Assessing the State of Televised Debate Styles in Latin America

This study assesses the state of political candidates’ verbal and nonverbal expressive behaviors during televised debates in four Latin American countries (Mexico, Colombia, Peru, and Argentina), a region largely overlooked in debate studies where uncivil behavior driven by populism and related ideologies is on the rise. Utilizing human annotation for both descriptive analysis and training and validation of machine learning classifiers, we analyzed 900 ten-second segments from recent debates in each country to detect behaviors associated with a populist style, including adversarial rhetoric and verbal attacks, an aggressive or threatening tone, angry facial expressions, and defiant gestures. We also identified agentic behaviors characterized by assertive, dynamic, and physically energetic actions, lacking the aggressive or threatening markers. Findings reveal distinct communication styles across countries coalescing around a core tendency. While aggressive and adversarial behaviors were anticipated, the predominant style observed was agentic, characterized by assertive and dynamic gestures absent displays of anger, defiance, or hostility. Agentic behavior was common despite the diversity in debate formats and party systems. The study underscores the importance of multimodal analysis in understanding political communication cross-nationally and provides insights into the political styles and performative strategies employed by Latin American candidates in the countries observed.

Weakly Supervised Nuclei Segmentation with Point-Guided Attention and Self-Supervised Pseudo-Labeling.

Due to the labor-intensive manual annotations for nuclei segmentation, point-supervised segmentation based on nuclei coordinate supervision has gained recognition in recent years. Despite great progress, two challenges hinder the performance of weakly supervised nuclei segmentation methods: (1) The stable and effective segmentation of adjacent cell nuclei remains an unresolved challenge. (2) Existing approaches rely solely on initial pseudo-labels generated from point annotations for training, and inaccurate labels may lead the model to assimilate a considerable amount of noise information, thereby diminishing performance. To address these issues, we propose a method based on center-point prediction and pseudo-label updating for precise nuclei segmentation. First, we devise a Gaussian kernel mechanism that employs multi-scale Gaussian masks for multi-branch center-point prediction. The generated center points are utilized by the segmentation module to facilitate the effective separation of adjacent nuclei. Next, we introduce a point-guided attention mechanism that concentrates the segmentation module's attention around authentic point labels, reducing the noise impact caused by pseudo-labels. Finally, a label updating mechanism based on the exponential moving average (EMA) and k-means clustering is introduced to enhance the quality of pseudo-labels. The experimental results on three public datasets demonstrate that our approach has achieved state-of-the-art performance across multiple metrics. This method can significantly reduce annotation costs and reliance on clinical experts, facilitating large-scale dataset training and promoting the adoption of automated analysis in clinical applications.

A novel deep learning-based pipeline architecture for pulp stone detection on panoramic radiographs.

Pulp stones are ectopic calcifications located in pulp tissue. The aim of this study is to introduce a novel method for detecting pulp stones on panoramic radiography images using a deep learning-based two-stage pipeline architecture. The first stage involved tooth localization with the YOLOv8 model, followed by pulp stone classification using ResNeXt. 375 panoramic images were included in this study, and a comprehensive set of evaluation metrics, including precision, recall, false-negative rate, false-positive rate, accuracy, and F1 score was employed to rigorously assess the performance of the proposed architecture. Despite the limited annotated training data, the proposed method achieved impressive results: an accuracy of 95.4%, precision of 97.1%, recall of 96.1%, false-negative rate of 3.9%, false-positive rate of 6.1%, and a F1 score of 96.6%, outperforming existing approaches in pulp stone detection. Unlike current studies, this approach adopted a more realistic scenario by utilizing a small dataset with few annotated samples, acknowledging the time-consuming and error-prone nature of expert labeling. The proposed system is particularly beneficial for dental students and newly graduated dentists who lack sufficient clinical experience, as it aids in the automatic detection of pulpal calcifications. To the best of our knowledge, this is the first study in the literature that propose a pipeline architecture to address the PS detection tasks on panoramic images.

Physics-Based Synthetic Data Model for Automated Segmentation in Catalysis Microscopy.

In catalysis research, the amount of microscopy data acquired when imaging dynamic processes is often too much for nonautomated quantitative analysis. Developing machine learned segmentation models is challenged by the requirement of high-quality annotated training data. We thus substitute expert-annotated data with a physics-based sequential synthetic data model. We study environmental scanning electron microscopy (ESEM) data collected from isopropanol oxidation to acetone over cobalt oxide as an example. Upon applying a temperature program during the reaction a phase transition occurs, reducing the catalyst selectivity toward acetone. This is accompanied on the micrometer ESEM scale by the formation of cracks between the pores of the catalyst surface. We aim to generate synthetic data to train a neural network capable of semantic segmentation (pixel-wise labeling) of this ESEM data. This analysis will lead to insights into this phase transition. To generate synthetic data that approximates this transition, our algorithm composes the ESEM images of the room-temperature catalyst with dynamically evolving synthetic cracks satisfying physical construction principles, gathered from qualitative knowledge accessible in the ESEM data. We mimic the surface crack growth propagation along surface paths, avoiding close vicinity to nearby pores. This physics-based approach results in a lowered rate of false positives compared to a random approach.

Validation and accuracy of the Hyfe cough monitoring system: a multicenter clinical study

Background: The ability to passively and continuously monitor coughing for prolonged periods of time would significantly improve cough management and research. To date there is no automated clinically validated cough monitor that can be routinely used in clinical care and research. Here we describe the validation of such an automated cough monitor. Methods: This multicenter observational study compared the results of the Hyfe CoughMonitor wrist-worn device with manually counted coughs in subjects with a variety of etiologies as they went about their usual daily activities. We collected 24 h of continuous sounds from subjects while they simultaneously wore a CoughMonitor and an audio recorder. Coughs were labelled by multiple trained annotators who listened to the continuous audio recordings using validated methodology. The time stamps of these human-detected coughs were compared to those of the CoughMonitor to determine the system’s overall performance using event-to-event and hourly rate correlation analyses. Results: Over the 546 h monitored, 4,454 cough events were recorded; The overall sensitivity was 90.4% (95% CI of 88.3–92.2%). The overall false positive rate was 1.03 false positives per hour (95% CI of 0.84 to 1.24). The overall correlation between manual and CoughMonitor measured hourly coughing was high (Pearson correlation coefficient of 0.99). Two case studies of long-term monitoring of patients with chronic cough are presented. Conclusion: The present analysis of cough events demonstrated that the Hyfe CoughMonitor accurately reflects them with a high sensitivity and a low false positive rate. Future studies should focus on its potential role in the management of patients with cough in clinical practice.Registration Clinicaltrials.gov, NCT05723159.

Preserving Fairness in AI under Domain Shift

Existing algorithms for ensuring fairness in AI use a single-shot training strategy, where an AI model is trained on an annotated training dataset with sensitive attributes and then fielded for utilization. This training strategy is effective in problems with stationary distributions, where both the training and testing data are drawn from the same distribution. However, it is vulnerable with respect to distributional shifts in the input space that may occur after the initial training phase. As a result, the time-dependent nature of data can introduce biases and performance degradation into the model predictions, even if the model is initially fair. Model retraining from scratch using a new annotated dataset is a naive solution that is expensive and time-consuming. We develop an algorithm to adapt a fair model to remain fair and generalizable under domain shift using solely new unannotated data points. We recast this learning setting as an unsupervised domain adaptation (UDA) problem. Our algorithm is based on updating the model such that the internal representation of data remains unbiased despite distributional shifts in the input space. We provide empirical validation on three common fairness datasets to show that the challenge exists in practical setting and to demonstrate the effectiveness of our algorithm.

Extreme R-CNN: Few-Shot Object Detection via Sample Synthesis and Knowledge Distillation.

Traditional object detectors require extensive instance-level annotations for training. Conversely, few-shot object detectors, which are generally fine-tuned using limited data from unknown classes, tend to show biases toward base categories and are susceptible to variations within these unknown samples. To mitigate these challenges, we introduce a Two-Stage Fine-Tuning Approach (TFA) named Extreme R-CNN, designed to operate effectively with extremely limited original samples through the integration of sample synthesis and knowledge distillation. Our approach involves synthesizing new training examples via instance clipping and employing various data-augmentation techniques. We enhance the Faster R-CNN architecture by decoupling the regression and classification components of the Region of Interest (RoI), allowing synthetic samples to train the classification head independently of the object-localization process. Comprehensive evaluations on the Microsoft COCO and PASCAL VOC datasets demonstrate significant improvements over baseline methods. Specifically, on the PASCAL VOC dataset, the average precision for novel categories is enhanced by up to 15 percent, while on the more complex Microsoft COCO benchmark it is enhanced by up to 6.1 percent. Remarkably, in the 1-shot scenario, the AP50 of our model exceeds that of the baseline model in the 10-shot setting within the PASCAL VOC dataset, confirming the efficacy of our proposed method.

Weakly-Supervised Hair SEM Microscope Image Segmentation Using a Priori Structure Information

The analysis of microscopic images of hair has a wide range of applications in the domains of cosmetics, healthcare and forensics. The segmentation of the hair represents the initial step of automatic large-scale quantitative analysis of microscopic images of hair. It ensures that subsequent quantitative measurements are performed in an appropriate area corresponding to the hair, avoiding artifacts near the boundaries. This process can be time-consuming, tedious and susceptible to subjective errors when conducted by a human operator. Deep learning methods represent a promising solution; however, obtaining pixel-level accurate masks is a costly process. This paper presents a novel weakly supervised pipeline for the segmentation of hair SEM (Scanning Electron Microscope) microscopic images, which requires only simple image-level annotations for training. The proposed method incorporates the Radon transform, the Sobel operator and a novel Boundary Discrimination Module (BD-module) for the estimation of the presence of boundaries. The proposed pipeline was evaluated on a recently collected hair SEM dataset (429 images). Furthermore, it is benchmarked with methods including Unet and Segment Anything Model (SAM). The results demonstrated a mean Hausdorff Distance improvement of over 30% and a standard deviation improvement of over 50% in comparison with the Unet and SAM. Moreover, we proposed additional refinement modules to address boundary nonlinear cases and conducted Grad-CAM analysis to enhance the interpretability of the BD-module. Additionally, we proposed a novel quality estimation metric based on gradient map for self-quality assessment. The SEM hair dataset is accessible to the research community in an open-source format.

CLIPMulti: Explore the performance of multimodal enhanced CLIP for zero-shot text classification

CLIPMulti: Explore the performance of multimodal enhanced CLIP for zero-shot text classification

Deep Learning-Based In Situ Micrograph Synthesis and Augmentation for Crystallization Process Image Analysis

Deep learning-based in situ imaging and analysis for crystallization process are essential for optimizing product qualities, reducing experimental costs through real-time monitoring, and controlling the process. However, large and high-quality annotated datasets are required to train accurate models, which are time consuming. Therefore, we proposed a novel methodology that applied image synthesis neural networks to generate virtual information-rich images, enabling efficient and rapid dataset expansion while simultaneously reducing annotation costs. Experiments were conducted on the L-alanine crystallization process to obtain process images and to validate the proposed workflow. The proposed method, aided by interpolation augmentation and data warping augmentation to enhance data richness, utilized only 25% of the training annotations, consistently segmenting crystallization process images comparable to those models utilizing 100% of the training data annotations, achieving an average precision of nearly 98%. Additionally, based on the analysis of Kullback–Leibler divergence, the proposed method demonstrated excellent performance in extracting in situ information regarding aspect ratios and crystal size distributions during the crystallization process. Moreover, its ability to leverage expert labels with a four-fold enhanced efficiency holds great potential for advancing various applications in crystallization processes.

Identifying retinal pigment epithelium cells in adaptive optics-optical coherence tomography images with partial annotations and superhuman accuracy.

Retinal pigment epithelium (RPE) cells are essential for normal retinal function. Morphological defects in these cells are associated with a number of retinal neurodegenerative diseases. Owing to the cellular resolution and depth-sectioning capabilities, individual RPE cells can be visualized in vivo with adaptive optics-optical coherence tomography (AO-OCT). Rapid, cost-efficient, and objective quantification of the RPE mosaic's structural properties necessitates the development of an automated cell segmentation algorithm. This paper presents a deep learning-based method with partial annotation training for detecting RPE cells in AO-OCT images with accuracy better than human performance. We have made the code, imaging datasets, and the manual expert labels available online.

Few-shot agricultural pest recognition based on multimodal masked autoencoder

Few-shot agricultural pest recognition based on multimodal masked autoencoder

PCLC‐Net: Point Cloud Completion in Arbitrary Poses with Learnable Canonical Space

AbstractRecovering the complete structure from partial point clouds in arbitrary poses is challenging. Recently, many efforts have been made to address this problem by developing SO(3)‐equivariant completion networks or aligning the partial point clouds with a predefined canonical space before completion. However, these approaches are limited to random rotations only or demand costly pose annotation for model training. In this paper, we present a novel Network for Point cloud Completion with Learnable Canonical space (PCLC‐Net) to reduce the need for pose annotations and extract SE(3)‐invariant geometry features to improve the completion quality in arbitrary poses. Without pose annotations, our PCLC‐Net utilizes self‐supervised pose estimation to align the input partial point clouds to a canonical space that is learnable for an object category and subsequently performs shape completion in the learned canonical space. Our PCLC‐Net can complete partial point clouds with arbitrary SE(3) poses without requiring pose annotations for supervision. Our PCLC‐Net achieves state‐of‐the‐art results on shape completion with arbitrary SE(3) poses on both synthetic and real scanned data. To the best of our knowledge, our method is the first to achieve shape completion in arbitrary poses without pose annotations during network training.

Tree hierarchical deep convolutional neural network optimized with sheep flock optimization algorithm for sentiment classification of Twitter data

ABSTRACT The increasing volume of online reviews and tweets poses significant challenges for sentiment classification because of the difficulty in obtaining annotated training data. This paper aims to enhance sentiment classification of Twitter data by developing a robust model that improves classification accuracy and computational efficiency. The proposed method named Tree Hierarchical Deep Convolutional Neural Network optimized with Sheep Flock Optimization Algorithm for Sentiment Classification of Twitter Data (SCTD-THDCNN-SFOA) utilizes the Stanford Sentiment Treebank dataset. The process begins with pre-processing steps including Tokenization, Stop words Elimination, Filtering, Hashtag Removal, and Multiword Grouping. The Gray Level Co-occurrence Matrix Window Adaptive Algorithm is employed to extract features, such as emoticon counts, punctuation counts, gazetteer word existence, n-grams, and part of speech tags. These features are selected using Entropy–Kurtosis-based Feature Selection approach. Finally, the Tree Hierarchical Deep Convolutional Neural Network enhanced by the Sheep Flock Optimization Algorithm is used to categorize the Twitter data as positive, negative, and neutral sentiments. The proposed SCTD-THDCNN-SFOA method demonstrates superior performance, achieving higher accuracy and lesser computation time than the existing models, respectively. The SCTD-THDCNN-SFOA framework significantly improves the accuracy and efficiency of sentiment classification for Twitter data.

A cross-domain transfer learning model for author name disambiguation on heterogeneous graph with pretrained language model

A cross-domain transfer learning model for author name disambiguation on heterogeneous graph with pretrained language model

Evaluation of segmentation and semi‐supervised learning for spotted spurge recognition in bermudagrass turf

AbstractDiverse datasets are crucial for training machine learning‐based weed recognition models. However, annotating (i.e., labeling) images can be laborious and time‐consuming. Choice of annotation method and training approach not only affects the overall model effectiveness but also its minimum training data requirements and development pace. Segmentation and semi‐supervised learning (SSL) may offer performance or training enhancements. This study evaluated (1) segmentation against object detection in spotted spurge [Chamaesyce maculata (L.) Small] recognition in Latitude 36 hybrid bermudagrass [Cynodon dactylon (L.) Pers. × C. transvaalensis Burtt‐Davy] maintained as a golf course fairway and (2) the potential of two‐step SSL‐based training procedure using data labeled both manually and automatically with a pre‐trained model to expedite the model development. The architecture used across this research was You Only Look Once version 8 (YOLOv8) employing nano, small, and medium variants. All models were trained with a dataset restricted to 1200 training and 300 validation images. Both annotation methods resulted in adequate spotted spurge identification, as evidenced with >0.60 (>0.50 acceptability threshold) medium average precision at intersection over union threshold of 0.50 (mAP@50). Although the difference was minimal, object detection performed superior to segmentation. The two‐step training effectively accelerated image annotation while preserving or improving (with 8:4 and 6:6 splits between manually and auto‐labeled data) object detection performance. Segmentation tolerated only 10:2 split and exhibited increased sensitivity to declining dataset quality proportionally to the increase in auto‐labeled images in the final training dataset. Findings show a two‐step SSL‐based training procedure expedites annotation, enhancing model development efficiency.

Precision HER2: a comprehensive AI system for accurate and consistent evaluation of HER2 expression in invasive breast Cancer

BackgroundWith the development of novel anti-HER2 targeted drugs, such as ADCs, it has become increasingly important to accurately interpret HER2 expression in breast cancer. Previous studies have demonstrated high intra-observer and inter-observer variabilities in evaluating HER2 staining by human eyes. There exists a strong requirement to develop artificial intelligence (AI) systems to achieve high-precision HER2 expression scoring for better clinical therapy.MethodsIn the present study, we collected breast cancer tissue samples and stained consecutive sections with anti-Calponin and anti-HER2 antibodies. High-quality digital images were selected from immunohistochemical slides and interpreted as HER2 3+, 2+, 1+, and 0. AI models were trained and assessed using annotated training and testing sets. The AI model was trained to automatically identify ductal carcinoma in situ (DCIS) by Calponin staining and myoepithelial annotation and filter out DCIS components in HER2-stained slides using image-overlapping techniques. Furthermore, we organized two-phase validation studies. In phase one, pathologists interpreted 112 HER2 whole-slide images (WSIs) without AI assistance, whereas in phase two, pathologists read the same slides using the AI system after a washing period of 2 weeks.ResultsOur AI model greatly improved the accuracy of reading (0.902 vs. 0.710). The number of HER2 1 + patients misdiagnosed as HER2 0 was significantly reduced (32/279 vs. 65/279), and they benefitted from ADC drugs. In addition, the AI algorithm improved the intra-group consistency of HER2 readings by pathologists with different years of experience (intra-class correlation coefficient [ICC]: 0.872–0.926 vs. 0.818–0.908), with the improvement most pronounced among junior pathologists (0.885 vs. 0.818).ConclusionsWe proposed a high-precision AI system to identify and filter out DCIS components and automatically evaluate HER2 expression in invasive breast cancer.