Deep Learning Detection Research Articles

LCGSC-YOLO: a lightweight apple leaf diseases detection method based on LCNet and GSConv module under YOLO framework

IntroductionIn response to the current mainstream deep learning detection methods with a large number of learned parameters and the complexity of apple leaf disease scenarios, the paper proposes a lightweight method and names it LCGSC-YOLO. This method is based on the LCNet(A Lightweight CPU Convolutional Neural Network) and GSConv(Group Shuffle Convolution) module modified YOLO(You Only Look Once) framework.MethodsFirstly, the lightweight LCNet is utilized to reconstruct the backbone network, with the purpose of reducing the number of parameters and computations of the model. Secondly, the GSConv module and the VOVGSCSP (Slim-neck by GSConv) module are introduced in the neck network, which makes it possible to minimize the number of model parameters and computations while guaranteeing the fusion capability among the different feature layers. Finally, coordinate attention is embedded in the tail of the backbone and after each VOVGSCSP module to improve the problem of detection accuracy degradation issue caused by model lightweighting.ResultsThe experimental results show the LCGSC-YOLO can achieve an excellent detection performance with mean average precision of 95.5% and detection speed of 53 frames per second (FPS) on the mixed datasets of Plant Pathology 2021 (FGVC8) and AppleLeaf9.DiscussionThe number of parameters and Floating Point Operations (FLOPs) of the LCGSC-YOLO are much less thanother related comparative experimental algorithms.

Research on Deep Learning Detection Model for Pedestrian Objects in Complex Scenes Based on Improved YOLOv7

Objective: Pedestrian detection is very important for the environment perception and safety action of intelligent robots and autonomous driving, and is the key to ensuring the safe action of intelligent robots and auto assisted driving. Methods: In response to the characteristics of pedestrian objects occupying a small image area, diverse poses, complex scenes and severe occlusion, this paper proposes an improved pedestrian object detection method based on the YOLOv7 model, which adopts the Convolutional Block Attention Module (CBAM) attention mechanism and Deformable ConvNets v2 (DCNv2) in the two Efficient Layer Aggregation Network (ELAN) modules of the backbone feature extraction network. In addition, the detection head is replaced with a Dynamic Head (DyHead) detector head with an attention mechanism; unnecessary background information around the pedestrian object is also effectively excluded, making the model learn more concentrated feature representations. Results: Compared with the original model, the log-average miss rate of the improved YOLOv7 model is significantly reduced in both the Citypersons dataset and the INRIA dataset. Conclusions: The improved YOLOv7 model proposed in this paper achieved good performance improvement in different pedestrian detection problems. The research in this paper has important reference significance for pedestrian detection in complex scenes such as small, occluded and overlapping objects.

Deep learning detection of subclinical cardiovascular dysfunction in type 2 diabetes through retinal photography

Abstract Background Type 2 diabetes (T2D) is associated with excess atherosclerotic coronary disease and heart failure risk, with several cardiac abnormalities described in asymptomatic patients. Routine digital retinal photography performed for screening of diabetic retinopathy (DR) has the potential to identify patients with underlying cardiovascular disease (CVD). We aimed to determine whether features extracted from digital retinal photographs could detect subclinical cardiovascular dysfunction in asymptomatic people with T2D. Methods We prospectively enrolled adults with T2D and no history or symptoms of CVD to undergo extensive phenotyping with multiparametric stress perfusion cardiac MRI and CT coronary artery calcium scoring. Digital retinal photographs acquired for routine diabetic eye screening and performed within one year of cardiac evaluation were interrogated for retinopathy grading and using four deep learning (DL) models (Xception, Inception, EfficientNet, and MobileNet) to capture complex patterns within the retinal vasculature. Each subject had four retinal photographs (two of each eye) for DL analysis. Cardiac image analysis was performed blinded to participant details and independent of retinal images. Subjects with and without DR were compared. Participants were divided into training (80%) and validation datasets (20%) prior to DL analysis. DL models employed to predict from retinal images: presence or absence of coronary artery calcium on non-contrast CT, myocardial perfusion reserve <2.5, left ventricular (LV) global longitudinal strain <16%, absolute stress and rest myocardial blood flood (mL/g/min) used as continuous variables. AUC was used to assess the ability of DL models to discriminate between those without and without subclinical CVD. Results An initial 254 subjects (comprising 1,016 retinal photographs) were included in the analysis: 212 (83%) with no DR, 42 (17%) with grade 1 (mild background) DR. Key cardiac imaging parameters in subjects with and without DR are displayed in Table 1. The groups were well matched for age, sex, body mass index and blood pressure. Overall, those with grade 1 DR had higher atheroma burden, evidence of increased LV filling pressures (E/e’) and poorer systolic function, but no differences in stress and rest myocardial blood flow or myocardial perfusion reserve. However, all four DL models failed to achieve sufficient discrimination of any of the evaluated cardiac imaging parameters, with the majority achieving AUC scores below 50% for all cardiac parameters. Conclusion Asymptomatic adults with grade 1 DR exhibit more subclinical atherosclerosis and cardiovascular dysfunction than those without DR. However, DL transfer models could not discriminate between subjects with and without cardiovascular disease. Further work is needed to better define whether retinal screening may be used to identify those people with T2D most at risk of underlying cardiovascular disease.Subjects demographics

Automated Car Damage Assessment Using Computer Vision: Insurance Company Use Case

Automated car damage detection using computer vision techniques has been studied using several datasets, but real cases for insurance companies are usually dependent on private methods and datasets. Furthermore, there are no metrics or standardized processes that describe the situation in which the company analyzes the customer’s images, the models used for the inference, and the results. We perform extensive experiments to show that our proposal, an ensemble of 10 deep learning detectors based on YOLOv5, improves the state-of-the-art not only in terms of typical metrics but also in terms of inference speed, allowing scalability to thousands of instances per minute. A comparison with YOLOv8 is carried out, showing the differences between both ensembles. Furthermore, a dataset called TartesiaDS, labeled under the supervision of professional appraisers from insurance companies, is available to the community for evaluation of future proposals.

Radio-opaque contrast agents for liver cancer targeting with KIM during radiation therapy (ROCK-RT): an observational feasibility study.

This observational study aims to establish the feasibility of using x-ray images of radio-opaque chemoembolisation deposits in patients as a method for real-time image-guided radiation therapy of hepatocellular carcinoma. This study will recruit 50 hepatocellular carcinoma patients who have had or will have stereotactic ablative radiation therapy and have had transarterial chemoembolisation with a radio-opaque agent. X-ray and computed tomography images of the patients will be analysed retrospectively. Additionally, a deep learning method for real-time motion tracking will be developed. We hypothesise that: (i) deep learning software can be developed that will successfully track the contrast agent mass on two thirds of cone beam computed tomography (CBCT) projection and intra-treatment images (ii), the mean and standard deviation (mm) difference in the location of the mass between ground truth and deep learning detection are ≤ 2mm and ≤ 3mm respectively and (iii) statistical modelling of study data will predict tracking success in 85% of trial participants. Developing a real-time tracking method will enable increased targeting accuracy, without the need for additional invasive procedures to implant fiducial markers. Registered to ClinicalTrials.gov (NCT05169177) 12th October 2021.

Deep Learning Detection of Hand Motion During Microvascular Anastomosis Simulations Performed by Expert Cerebrovascular Neurosurgeons

Deep learning enables precise hand tracking without the need for physical sensors, allowing for unsupervised quantitative evaluation of surgical motion and tasks. We quantitatively assessed the hand motions of experienced cerebrovascular neurosurgeons during simulated microvascular anastomosis using deep learning. We explored the extent to which surgical motion data differed among experts. A deep learning detection system tracked 21 landmarks corresponding to digit joints and the wrist on each hand of 5 expert cerebrovascular neurosurgeons. Tracking data for each surgeon were analyzed over long and short time intervals to examine gross movements and micromovements, respectively. Quantitative algorithms assessed the economy and flow of motion by calculating mean movement distances from the baseline median landmark coordinates and median times between sutures, respectively. Tracking data correlated with specific surgical actions observed in microanastomosis video analysis. Economy of motion during suturing was calculated as 19, 26, 29, 27, and 28 pixels for surgeons 1, 2, 3, 4, and 5, respectively. Flow of motion during microanastomosis was 31.96, 29.40, 28.90, 7.37, and 47.21seconds for surgeons 1, 2, 3, 4, and 5, respectively. Hand tracking data showed similarities among experts, with low movements from baseline, minimal excess motion, and rhythmic suturing patterns. The data revealed unique patterns related to each expert's habits and techniques. The results showed that surgical motion can be correlated with hand motion and assessed using mathematical algorithms. We also demonstrated the feasibility and potential of deep learning-based motion detection to enhance surgical training.

Deep Learning-Based Biomimetic Identification Method for Mask Wearing Standardization.

Deep learning technology can automatically learn features from large amounts of data, with powerful feature extraction and pattern recognition capabilities, thereby improving the accuracy and efficiency of object detection. [The objective of this study]: In order to improve the accuracy and speed of mask wearing deep learning detection models in the post pandemic era, the [Problem this study aimed to resolve] was based on the fact that no research work has been reported on standardized detection models for mask wearing with detecting nose targets specially. [The topic and method of this study]: A mask wearing normalization detection model (towards the wearing style exposing the nose to outside, which is the most obvious characteristic of non-normalized style) based on improved YOLOv5s (You Only Look Once v5s is an object detection network model) was proposed. [The improved method of the proposed model]: The improvement design work of the detection model mainly includes (1) the BottleneckCSP (abbreviation of Bottleneck Cross Stage Partial) module was improved to a BottleneckCSP-MASK (abbreviation of Bottleneck Cross Stage Partial-MASK) module, which was utilized to replace the BottleneckCSP module in the backbone architecture of the original YOLOv5s model, which reduced the weight parameters' number of the YOLOv5s model while ensuring the feature extraction effect of the bonding fusion module. (2) An SE module was inserted into the proposed improved model, and the bonding fusion layer in the original YOLOv5s model was improved for better extraction of the features of mask and nose targets. [Results and validation]: The experimental results indicated that, towards different people and complex backgrounds, the proposed mask wearing normalization detection model can effectively detect whether people are wearing masks and whether they are wearing masks in a normalized manner. The overall detection accuracy was 99.3% and the average detection speed was 0.014 s/pic. Contrasted with original YOLOv5s, v5m, and v5l models, the detection results for two types of target objects on the test set indicated that the mAP of the improved model increased by 0.5%, 0.49%, and 0.52%, respectively, and the size of the proposed model compressed by 10% compared to original v5s model. The designed model can achieve precise identification for mask wearing behaviors of people, including not wearing a mask, normalized wearing, and wearing a mask non-normalized.

SCINet: A Segmentation and Classification Interaction CNN Method for Arteriosclerotic Retinopathy Grading.

As a common disease, cardiovascular and cerebrovascular diseases pose a great harm threat to human wellness. Even using advanced and comprehensive treatment methods, there is still a high mortality rate. Arteriosclerosis, as an important factor reflecting the severity of cardiovascular and cerebrovascular diseases, is imperative to detect the arteriosclerotic retinopathy. However, the detection of arteriosclerosis retinopathy requires expensive and time-consuming manual evaluation, while end-to-end deep learning detection methods also need interpretable design to high light task-related features. Considering the importance of automatic arteriosclerotic retinopathy grading, we propose a segmentation and classification interaction network (SCINet). We propose a segmentation and classification interaction architecture for grading arteriosclerotic retinopathy. After IterNet is used to segment retinal vessel from original fundus images, the backbone feature extractor roughly extracts features from the segmented and original fundus arteriosclerosis images and further enhances them through the vessel aware module. The last classifier module generates fundus arteriosclerosis grading results. Specifically, the vessel aware module is designed to highlight the important areal vessel features segmented from original images by attention mechanism, thereby achieving information interaction. The attention mechanism selectively learns the vessel features of segmentation region information under the proposed interactive architecture, which leads to reweighting the extracted features and enhances significant feature information. Extensive experiments have confirmed the effect of our model. SCINet has the best performance on the task of arteriosclerotic retinopathy grading. Additionally, the CNN method is scalable to similar tasks by incorporating segmented images as auxiliary information.

Histopathologic Differential Diagnosis and Estrogen Receptor/Progesterone Receptor Immunohistochemical Evaluation of Breast Carcinoma Using a Deep Learning–Based Artificial Intelligence Architecture

In breast carcinoma, invasive ductal carcinoma (IDC) is the most common histopathological subtype, and ductal carcinoma in situ (DCIS) is a precursor of IDC. They are often concomitant. The immunohistochemical staining of estrogen receptor (ER)/progesterone receptor (PR) in IDC/DCIS on whole-slide histopathological images (WSIs) can predict the prognosis of patients. However, the inter-observer variability among pathologists in reading WSIs is inevitable. Thus, artificial intelligence (AI) technology is crucial. Herein, IDC/DCIS detection was conducted by deep learning approach, including Faster R-CNN, RetinaNet, SSD300, YOLOv3, YOLOv5, YOLOv7, YOLOv8, and Swin transformer. Their performance was estimated by mean average precision (mAP) values. Cell recognition and counting were performed using AI technology to evaluate the intensity and proportion of ER/PR-immunostained cancer cells in IDC/DCIS. A three-round ring study (RS) was conducted to assess WSIs. A database for modelling the underlying probability distribution of a dataset with labels was established. YOLOv8 exhibits the highest detection performance with an mAP@0.5 of 0.944 and an mAP@0.5-0.95 of 0.790. With the assistance of YOLOv8, the scoring concordance across all pathologists was boosted to excellent in RS3 (0.970) from moderate in RS1 (0.724) and good in RS2 (0.812). Deep learning detection can be applied in clinicopathological field. To facilitate the histopathological diagnosis of IDC/DCIS and immunostaining scoring of ER/PR, a novel AI architecture and well-organized dataset were developed.

NEFELI: A deep-learning detection and tracking pipeline for enhancing autonomy in advanced air mobility

Efficient detection and accurate collision estimation for non-cooperative aerial vehicles are crucial for the realization of fully autonomous aircraft and Advanced Air Mobility (AAM). This paper introduces NEFELI, a machine learning software utilizing optical sensors to detect and track non-cooperative aerial vehicles. NEFELI's detector employs an enhanced YOLOv5-large model, strengthened with a sliced inference step to enhance detection capabilities for distant, diminutive objects. Furthermore, NEFELI introduces several innovations in its tracking component. A key advancement lies in the creation and utilization of the first large-scale re-identification (Re-ID) dataset of aerial objects. This dataset is used to train the deep learning appearance (Re-ID) model of the tracking module and integrates appearance information into the detection and tracking pipeline, resulting in more robust and reliable tracking performance. Moreover, the tracking model combines the deep learning appearance model with a Kalman Filter-based motion model to address the challenge of precisely tracking distant aerial objects. Notably, an extensive comparative analysis that was conducted showed that NEFELI outperforms state-of-the-art detection and tracking models in terms of Higher Order Tracking Accuracy (HOTA) metric, ID switches, and Association Accuracy (AssA) by a wide margin. A crucial aspect of this work is NEFELI's software architecture design, which enables efficient implementation on a low SWaP (Size, Weight, and Power) edge Graphic Processing Unit (GPU). To further showcase NEFELI's generalization capabilities and edge implementation performance, real-world flight experiments with small UAVs were carried out. The experimental results demonstrate NEFELI's ability to detect and track small UAVs at distances of up to 145 m in real-time speed of 6.7 fps.

Application of Deep Learning Model and Machine Learning Detection Algorithm for Gastric Cancer Patients Undergoing Chemotherapy

This work focused on exploring the application value of machine learning detection (MLD) algorithm and evidence-based nursing (EBN) in the chemotherapy (CHET) for gastric cancer (GC) patients. 100 GC patients who were treated in the Guang’an Traditional Chinese Medicine Hospital and needed postoperative CHET were recruited and randomly assigned to experimental (Exp) and control (Ctrl) groups, each including 50 patients. All participants received adjuvant CHET after gastrectomy. During CHET, participants in the Ctrl group were given routine nursing, while the experimental were given EBN in addition to routine nursing. Differences in self-rating anxiety scale (SAS), self-rating depress scale (SDS), QLQ-C30 life core questionnaire, and adverse reaction (AR) evaluation criteria were compared for participants in different groups after nursing. At the same time, all patients underwent computed tomography (CT) examination and all images were detected by MLD algorithm. After intervention, the SAS and SDS scores of patients in the Exp and Ctrl groups were 26.7±5.3 versus 33.6±6.61 and 30.07±5.58 versus 36.11±8.83, respectively. The total health status (THS) score of patients was 5.59±1.17 in Exp group and 4.53±0.96 in Ctrl group, showing P < 0.05. After intervention, great differences were observed in nausea/vomiting, decreased white blood cells (WBC), decreased haemoglobin (Hb), peripheral nerve paraesthesia (PNP), muscle and joint pain (MJP), hair loss, and other indicators between patients received EBN and routine nursing methods (P < 0.05). The MLD algorithm and EBN were of high application value in the nursing of CHET treatment for GC patients.

Integrated Circuit Packaging Defect Analysis and Deep Learning Detection Method

Integrated Circuit Packaging Defect Analysis and Deep Learning Detection Method

Deep learning models for regional phase detection on seismic stations in Northern Europe and the European Arctic

SUMMARY Seismic phase detection and classification using deep learning is so far poorly investigated for regional events since most studies focus on local events and short time windows as the input to the detection models. To evaluate deep learning on regional seismic records, we create a data set of events in Northern Europe and the European Arctic. This data set consists of about 151 000 three component event waveforms and corresponding phase arrival picks at stations in mainland Norway, Finland and Svalbard. We train several state-of-the-art and one newly developed deep learning model on this data set to pick P- and S-wave arrivals. The new method modifies the popular PhaseNet model with new convolutional blocks including transformers. This yields more accurate predictions on the long input time windows associated with regional events. Evaluated on event records not used for training, our new method improves the performance of the current state-of-the-art methods when it comes to recall, precision and pick time residuals. Finally, we test our new model for continuous mode processing on 4 d of single-station data from the ARCES array. Results show that our new method outperforms the existing array detector at ARCES. This opens up new opportunities to improve automatic array processing with deep learning detectors.

FDIA localization and classification detection in smart grids using multi-modal data and deep learning technique

False data injection attacks (FDIA) exploit the vulnerabilities of bad data detection in energy management systems to maliciously tamper with state estimation results, seriously jeopardizing the safe and reliable operation of power systems. In order to promptly detect FDIA, recent studies have used machine learning techniques to extract attack characteristics and detect FDIA based on changes in these characteristics. Current research predominantly focuses on detecting a single type of FDIA attack model and topology. However, the increasing diversification of FDIA construction methods and the variability of topological structures in power systems can reduce or even invalidate detection effectiveness. To cope with the abovementioned difficulties, this study introduces a multimodal deep learning detection model based on variational graph auto-encoders (VGAE), temporal convolutional networks (TCN), and gated recurrent units (GRU). The topological features of power systems are obtained through VGAE to adapt to the detection needs of various topological structures and performance enhancement. These features are then integrated with preprocessed measurement data via BDD to form multimodal data. Furthermore, this multimodal data is used to locate and classify FDIA with complete information, FDIA with incomplete information, and topology attack after applying a multi-label classification algorithm based on TCN-GRU for temporal feature extraction. Experiments carried out on the IEEE 14 and IEEE 118 bus systems show that the proposed method is robust and has high detection performance. The results indicate that the proposed detection method achieves AUC values that are, on average, 0.085, 0.145, and 0.04 higher than those of CNN, LSTM, and CNN-LSTM, respectively. Moreover, under various noise and attack intensities, recall, precision, F1 score, and RACC remain above 0.859, 0.877, 0.867, and 0.818, respectively, with a classification accuracy greater than 0.912. This study provides a unique perspective on detecting FDIA across various attack models and power system topologies.

Deep learning for detection and counting of Nephrops norvegicus from underwater videos

Abstract The Norway lobster (Nephrops norvegicus) is one of the most important fishery items for the EU blue economy. This paper describes a software architecture based on neural networks, designed to identify the presence of N. norvegicus and estimate the number of its individuals per square meter (i.e. stock density) in deep-sea (350–380 m depth) Fishery No-Take Zones of the northwestern Mediterranean. Inferencing models were obtained by training open-source networks with images obtained from frames partitioning of in submarine vehicle videos. Animal detections were also tracked in successive frames of video sequences to avoid biases in individual recounting, offering significant success and precision in detection and density estimations.

Self-supervised Domain Adaptation with Significance-Oriented Masking for Pelvic Organ Prolapse detection

Pelvic Organ Prolapse (POP) is a common disease in middle-aged and elderly women. The detection of POP is a challenging task, and using deep learning for detection has its practical significance. However, medical image detection tasks always face many problems, i.e., small sample size, data imbalance, and unobvious pathological characteristics. In this paper, we propose a new training framework, called self-supervised Domain Adaptation with Significance-Oriented Masking (DASOM), to address these problems and improve the performance of POP intelligent detection. DASOM includes a new pre-training process based on the masked image modeling task, and redesigns the masking strategy, bringing the local induction capability required for detection to the model. Meanwhile, we also adopt the data process method fitting the pelvic floor ultrasonic dataset to effectively solve the problem of data shortage and imbalance. Extensive experimental results and analysis confirm that the proposed method significantly improves the performance and reliability of POP detection.

Automatic deep learning detection of overhanging restorations in bitewing radiographs.

This study aimed to assess the effectiveness of deep convolutional neural network (CNN) algorithms for the detecting and segmentation of overhanging dental restorations in bitewing radiographs. A total of 1160 anonymized bitewing radiographs were used to progress the artificial intelligence (AI) system for the detection and segmentation of overhanging restorations. The data were then divided into three groups: 80% for training (930 images, 2399 labels), 10% for validation (115 images, 273 labels), and 10% for testing (115 images, 306 labels). A CNN model known as You Only Look Once (YOLOv5) was trained to detect overhanging restorations in bitewing radiographs. After utilizing the remaining 115 radiographs to evaluate the efficacy of the proposed CNN model, the accuracy, sensitivity, precision, F1 score, and area under the receiver operating characteristic curve (AUC) were computed. The model demonstrated a precision of 90.9%, a sensitivity of 85.3%, and an F1 score of 88.0%. Furthermore, the model achieved an AUC of 0.859 on the receiver operating characteristic (ROC) curve. The mean average precision (mAP) at an intersection over a union (IoU) threshold of 0.5 was notably high at 0.87. The findings suggest that deep CNN algorithms are highly effective in the detection and diagnosis of overhanging dental restorations in bitewing radiographs. The high levels of precision, sensitivity, and F1 score, along with the significant AUC and mAP values, underscore the potential of these advanced deep learning techniques in revolutionizing dental diagnostic procedures.

Enhancing automatic prediction of clinically significant prostate cancer with deep transfer learning 2.5-dimensional segmentation on bi-parametric magnetic resonance imaging (bp-MRI).

The aggressiveness of prostate cancer (PCa) is crucial in determining treatment method. The purpose of this study was to establish a 2.5-dimensional (2.5D) deep transfer learning (DTL) detection model for the automatic detection of clinically significant PCa (csPCa) based on bi-parametric magnetic resonance imaging (bp-MRI). A total of 231 patients, including 181 with csPCa and 50 with non-clinically significant PCa (non-csPCa), were enrolled. Stratified random sampling was then employed to divide all participants into a training set [185] and a test set [46]. The DTL model was obtained through image acquisition, image segmentation, and model construction. Finally, the diagnostic performance of the 2.5D and 2-dimensional (2D) models in predicting the aggressiveness of PCa was evaluated and compared using receiver operating characteristic (ROC) curves. DTL models based on 2D and 2.5D segmentation were established and validated to assess the aggressiveness of PCa. The results demonstrated that the diagnostic efficiency of the DTL model based on 2.5D was superior to that of the 2D model, regardless of whether in a single or combined sequence. Particularly, the 2.5D combined model outperformed other models in differentiating csPCa from non-csPCa. The area under the curve (AUC) values for the 2.5D combined model in the training and test sets were 0.960 and 0.949, respectively. Furthermore, the T2-weighted imaging (T2WI) model showed superiority over the apparent diffusion coefficient (ADC) model, but was not as effective as the combined model, whether based on 2.5D or 2D. A DTL model based on 2.5D segmentation was developed to automatically evaluate PCa aggressiveness on bp-MRI, improving the diagnostic performance of the 2D model. The results indicated that the continuous information between adjacent layers can enhance the detection rate of lesions and reduce the misjudgment rate based on the DTL model.

FP‐Flow: Feature pyramid flow model for fabric defect detection

AbstractFabric defect detection is a crucial aspect of fabric production. At present, deep learning detection methods mostly rely on supervised learning. To tackle this issue, this study proposes an unsupervised fabric defect detection approach based‐on normalising flow. The method only needs to train the mapping of the feature probability distribution of defect‐free samples to a Gaussian distribution. In the inference process, the location of defects can be determined by testing the distance between the probability distribution of image features and the estimated distribution. To adapt to the complex background and various minor defects of the fabric, a feature pyramid structure is adopted. Moreover, considering the gradient vanishing and network degradation caused by deep layers during training, a residual structure is incorporated into the model. Experimental results demonstrate that the feature pyramid flow model outperforms other methods in defect detection across multiple datasets, with an average score rate of 98.7% and 100% for pixel‐level area under the curve (AUC) receiver operating characteristic and image‐level AUC, respectively, compared to an average score rate of 91.1% and 85.4% for other methods.

Deep generative model of the distal tibial classic metaphyseal lesion in infants: assessment of synthetic images.

The classic metaphyseal lesion (CML) is a distinctive fracture highly specific to infant abuse. To increase the size and diversity of the training CML database for automated deep-learning detection of this fracture, we developed a mask conditional diffusion model (MaC-DM) to generate synthetic images with and without CMLs. To objectively and subjectively assess the synthetic radiographic images with and without CMLs generated by MaC-DM. For retrospective testing, we randomly chose 100 real images (50 normals and 50 with CMLs; 39 infants, male = 22, female = 17; mean age = 4.1 months; SD = 3.1 months) from an existing distal tibia dataset (177 normal, 73 with CMLs), and generated 100 synthetic distal tibia images via MaC-DM (50 normals and 50 with CMLs). These test images were shown to 3 blinded radiologists. In the first session, radiologists determined if the images were normal or had CMLs. In the second session, they determined if the images were real or synthetic. We analyzed the radiologists' interpretations and employed t-distributed stochastic neighbor embedding technique to analyze the data distribution of the test images. When presented with the 200 images (100 synthetic, 100 with CMLs), radiologists reliably and accurately diagnosed CMLs (kappa = 0.90, 95% CI = [0.88-0.92]; accuracy = 92%, 95% CI = [89-97]). However, they were inaccurate in differentiating between real and synthetic images (kappa = 0.05, 95% CI = [0.03-0.07]; accuracy = 53%, 95% CI = [49-59]). The t-distributed stochastic neighbor embedding analysis showed substantial differences in the data distribution between normal images and those with CMLs (area under the curve = 0.996, 95% CI = [0.992-1.000], P < .01), but minor differences between real and synthetic images (area under the curve = 0.566, 95% CI = [0.486-0.647], P = .11). Radiologists accurately diagnosed images with distal tibial CMLs but were unable to distinguish real from synthetically generated ones, indicating that our generative model could synthesize realistic images. Thus, MaC-DM holds promise as an effective strategy for data augmentation in training machine-learning models for diagnosis of distal tibial CMLs.