Small Imaging Research Articles

Matrix-assisted laser desorption/ionization matrix incorporation evaluation algorithm for improved peak coverage and signal-to-noise ratio in mass spectrometry imaging.

Despite decades of implementation, the selection of optimal sample preparation conditions for matrix-assisted laser desorption/ionization (MALDI) imaging is still ambiguous due to the lack of a universal and comprehensive evaluation methodology. Thus, numerous experiments with different matrix application conditions accompany a translation of the method to novel sample types and matrices. Mouse brain tissues were covered with 9-aminoacridine through sublimation, followed by recrystallization in vapors of 5% (v/v) methanol solution in water. The samples were analyzed by MALDI time-of-flight mass spectrometry, and the efficiency of lipid and small-molecule ionization was evaluated with different metrics. We first investigate the dependency of matrix density and recrystallization conditions on the thickness of an analyte-empty matrix layer to roughly evaluate the laser shot number required to obtain an intense signal with minimal noise. Then, we introduce metrics for the analysis of small imaging datasets (small sample regions) of model samples based on median quantity of peaks in spectra (medQP) and weighted median signal-to-noise ratio (wmSNR). The evaluation of small regions and taking median values for metrics help overcome the sample heterogeneity and allow for the simultaneous comparison of different acquisition parameters. Here, we propose a methodology based on gradual laser ablation of small regions of sample and further implementation of weighted signal-to-noise ratio to assess various matrix application conditions. The proposed approach helps reduce the number of test samples required to determine optimal sample preparation conditions and improve the overall quality of images.

Augmenting X-ray micro-CT data with MICP data for high resolution pore-scale simulations of flow properties of carbonate rocks

Pore-scale modeling is essential in understanding and predicting flow and transport properties of rocks. Generally, pore-scale modeling is dependent on imaging technologies such as Micro Computed Tomography (micro-CT), which provides visual confirmation into the pore microstructures of rocks at a representative scale. However, this technique is limited in the ability to provide high resolution images showing the pore-throats connecting pore bodies. Pore scale simulations of flow and transport properties of rocks are generally done on a single 3D pore microstructure image. As such, the simulated properties are only representative of the simulated pore-scale rock volume. These are the technological and computational limitations which we address here by using a stochastic pore-scale simulation approach. This approach consists of constructing hundreds of 3D pore microstructures of the same pore size distribution and overall porosity but different pore connectivity. The construction of the 3D pore microstructures incorporates the use of Mercury Injection Capillary Pressure (MICP) data to account for pore throat size distribution, and micro-CT images to account for pore body size distribution. The approach requires a small micro-CT image volume (7–19 mm3) to reveal key pore microstructural features that control flow and transport properties of highly heterogeneous rocks at the core-scale. Four carbonate rock samples were used to test the proposed approach. Permeability calculations from the introduced approach were validated by comparing laboratory measured permeability of rock cores and permeability estimated using five well-known core-scale empirical model equations. The results show that accounting for the stochastic connectivity of pores results in a probabilistic distribution of flow properties which can be used to upscale pore-scale simulated flow properties to the core-scale. The use of the introduced stochastic pore-scale simulation approach is more beneficial when there is a higher degree of heterogeneity in pore size distribution. This is shown to be the case with permeability and hydraulic tortuosity which are key controls of flow and transport processes in rocks.

Automated segmentation of large image datasets using artificial intelligence for microstructure characterisation and damage analysis

Many properties of commonly used materials are driven by their microstructure, which can be influenced by the composition and manufacturing processes. To optimise future materials, understanding the microstructure is critically important. Here, we present two novel approaches based on artificial intelligence that allow the segmentation of the phases of a microstructure for which simple numerical approaches, such as thresholding, are not applicable: One is based on the nnU-Net neural network, and the other on generative adversarial networks (GAN).Using scanning electron microscopy images collected from large areas (∼1 mm2) of dual-phase steels as a case study, we demonstrate how both methods effectively segment intricate microstructural details, including martensite, ferrite, and damage sites, for subsequent analysis.Either method shows substantial generalizability across a range of image sizes and conditions, including heat-treated microstructures with different phase configurations. The nnU-Net excels in mapping large image areas. Conversely, the GAN-based method performs reliably on smaller images, providing greater step-by-step control and flexibility over the segmentation process.This study highlights the benefits of segmented microstructural data for various purposes, such as calculating phase fractions, modelling material behaviour through finite element simulation, and conducting geometrical analyses of damage sites and the local properties of their surrounding microstructure.

Patch-based probabilistic identification of plant roots using convolutional neural networks

Recently, computer vision and artificial intelligence are being used as enabling technologies for plant phenotyping studies, since they allow the analysis of large amounts of data gathered by the sensors. Plant phenotyping studies can be devoted to the evaluation of complex plant traits either on the aerial part of the plant as well as on the underground part, to extract meaningful information about the growth, development, tolerance, or resistance of the plant itself. All plant traits should be evaluated automatically and quantitatively measured in a non-destructive way. This paper describes a novel approach for identifying plant roots from images of the root system architecture using a convolutional neural network (CNN) that operates on small image patches calculating the probability that the center point of the patch is a root pixel. The underlying idea is that the CNN model should embed as much information as possible about the variability of the patches that can show chaotic and heterogeneous backgrounds. Results on a real dataset demonstrate the feasibility of the proposed approach, as it overcomes the current state of the art.

Bundle Adjustment of Aerial Linear Pushbroom Hyperspectral Images with Sub-Pixel Accuracy

AbstractLinear pushbroom (LP) cameras are often used in airborne hyperspectral imaging (HSI). Orthoimages are generated for HSI analyses but these require accurate camera orientations from bundle adjustment. However, the limited image overlap leads to an over-parameterization in the bundle adjustment with six degrees of freedom per image exposure, i.e., LP image line. Feature-based matching based on salient key points in small image neighborhoods cannot be readily applied to LP image lines, since each LP image line is only a single pixel wide in one of the image dimensions. A naive mosaicing of consecutive LP image lines leads to unacceptably large errors owing to the relative camera motion within these mosaics, even for a stabilized camera. Thus, a new method that allows the use of established methods for feature-based matching from aerial LP image lines is presented, and observations are retrieved and used in the bundle adjustment. The examination of the spatial misregistration between spectral bands in HSI cameras, i.e., the chromatic aberration, is also examined. The method assumes a stabilized camera system with a processed global navigation satellite system aided inertial navigation system solution. The bundle adjustment is done by estimating trajectory corrections in time intervals and retrieving discrete trajectory estimates from cubic spline interpolation. An experiment was conducted to demonstrate the method. The chromatic aberration is shown to be of sub-pixel level in the LP HSI camera, and the resulting planimetric accuracy (normalized median absolute deviation) from the bundle adjustment is $$\sim 1/4$$ ∼ 1 / 4 of the ground sampling distance in each of the north and east components. The accurate estimates from the bundle adjustment are shown to be suitable for high-quality orthoimage generation.

Calibration method for binocular vision system with large field of view based on small target image splicing

In vision measurement, camera calibration has a significant impact on measurement precision. The classical target-based calibration methods require the target to occupy more than one-third of the field of view. A small-size target that does not meet the requirements results in poor calibration accuracy, while an appropriate large-size target is difficult to manufacture and inconvenient to operate. In view of the above problem, we propose a flexible and accurate calibration method based on small target image splicing to calibrate the binocular vision system with a large field of view. The spliced images and virtual large targets are constructed to extend the target size, providing better flexibility for calibration. Moreover, an optimization objective function integrating two constraints in the imaging plane and measurement space is presented to improve the calibration accuracy during the parameter optimization process. The simulation experiments and actual experiments are carried out to test the performance of the proposed method. The results demonstrate that the calibration accuracy of the proposed method using a small target is equivalent to that of Zhang’s method using a large target. Additionally, when using a same-size target, the parameter error of the proposed method is less than that of Zhang’s method, and the proposed method reduces the distance measurement error from 1.169 mm to 0.208 mm compared to Zhang’s method.

Multi-hop double-branch network for small sample hyperspectral image classification

ABSTRACT Neural networks have been extensively utilized in the classification of hyperspectral images. The study introduces a novel dual-branch structured network named MGACN for hyperspectral image classification under the condition of small sample sizes. This model a multi-hop graph network (MHGCN) and a graph attention convolutional network (GACNN). Graph Convolutional Network (GCN) integrates multi-hop mechanism to obtain and aggregate information between long-distance non-neighbor nodes. The Graph Attention Network (GAT) combines 2D-CNN by incorporating a noise reduction mechanism, as well as position and channel attention mechanisms. The classification accuracy reaches 98.43%, 99.02%, and 98.55% on the Indian Pines, Salinas, and Pavia University datasets, which is significantly better than that of the comparison model.

Retinal Ischemic Perivascular Lesion Reflects Cerebral Small Vessel Disease Burden in Single Subcortical Infarction.

Retinal ischemic perivascular lesions (RIPLs) are an indicator of ischemia in the middle retina. We aimed to determine the relationship between RIPLs and single subcortical infarction (SSI). We also investigated the differences in cerebral small vessel disease imaging burden between groups with and without RIPLs in SSI. In this case-control study, we enrolled 82 patients with SSI and 72 nonstroke controls. All participants underwent magnetic resonance imaging and swept-source optical coherence tomography/optical coherence tomography angiography. Small vessel disease markers such as lacunes, cerebral microbleeds, white matter hyperintensity, and perivascular spaces were rated on brain imaging. RIPLs were assessed via swept-source optical coherence tomography. Optical coherence tomography angiography was used to measure the superficial vascular complex and deep vascular complex of the retina. After adjusting for risk factors, the presence of RIPLs was significantly associated with SSI (odds ratio [OR], 1.506 [95% CI, 1.365-1.662], P<0.001). Eyes with RIPLs showed lower deep vascular complex density (P=0.035) compared with eyes without RIPLs in patients with SSI. After adjusting for vascular risk factors, the presence of RIPLs in patients with SSI was associated with an increased periventricular white matter hyperintensity burden (β=0.414 [95% CI, 0.181-0.647], P<0.001) and perivascular spaces-basal ganglia (β=0.296 [95% CI, 0.079-0.512], P=0.008). RIPLs are associated with SSI independent of underlying risk factors. The relationship between the presence of RIPLs and small vessel disease markers provides evidence that RIPLs might be an additional indicator of cerebral ischemic changes.

Bronze Anthropomorphic Figurines from the Sacrificial Pits of Sanxingdui

This work gives a description of bronze figurines from the second sacrificial pit of Sanxingdui and their classification according to the affinity to different categories of ritual bronzes whose details they initially were. The classification helped to attribute small anthropomorphic images to one of the, earlier isolated by the author, stages of the technological evolution of Sanxingdui art of bronze casting from the JK2 sacrificial pit. Therefore, the chronologically meaningful typology of bronze figurines from the second sacrificial pit was constructed. As a result, the following groups of figurines were isolated: 1) figurines sitting at the roots of the “spirit trees”, 2) figurines that were details of the so-called “sacred altar” (“altar of spirits”), 3) figurines holding two-horned flat sceptres in their hands, 4) figurines kneeling on pedestals. Figurines sitting at the roots of the “spirit trees” belong to the first and second stages of the technological evolution of bronzes from the JK2 sacrificial pit. The sets of three figurines sitting at the roots of the “spirit trees” no. 2 and no. 3 represent in each case only one personage, so it can be seen from every side of the tree. All the figurines from the “sacred altar” must be dated to the third stage of the technological evolution of bronzes from JK2 like the artifact being discussed. Analogies to figurines from the “sacred altar” can be found among jades from the urn burials of Xiaojiawuji cemetery belonging to the late Neolithic Shijiahe culture of Hubei province. Bronze figurines also give an opportunity to reconstruct garments of the ancient inhabitants of Sanxingdui.

A Lightweight Remote Sensing Small Target Image Detection Algorithm Based on Improved YOLOv8.

In response to the challenges posed by small objects in remote sensing images, such as low resolution, complex backgrounds, and severe occlusions, this paper proposes a lightweight improved model based on YOLOv8n. During the detection of small objects, the feature fusion part of the YOLOv8n algorithm retrieves relatively fewer features of small objects from the backbone network compared to large objects, resulting in low detection accuracy for small objects. To address this issue, firstly, this paper adds a dedicated small object detection layer in the feature fusion network to better integrate the features of small objects into the feature fusion part of the model. Secondly, the SSFF module is introduced to facilitate multi-scale feature fusion, enabling the model to capture more gradient paths and further improve accuracy while reducing model parameters. Finally, the HPANet structure is proposed, replacing the Path Aggregation Network with HPANet. Compared to the original YOLOv8n algorithm, the recognition accuracy of mAP@0.5 on the VisDrone data set and the AI-TOD data set has increased by 14.3% and 17.9%, respectively, while the recognition accuracy of mAP@0.5:0.95 has increased by 17.1% and 19.8%, respectively. The proposed method reduces the parameter count by 33% and the model size by 31.7% compared to the original model. Experimental results demonstrate that the proposed method can quickly and accurately identify small objects in complex backgrounds.

Heightened infiltration of immune effector cell types in the hippocampal regions in the APP/PS1 mouse model of Alzheimer’s disease.

Abstract Alzheimer's disease (AD) is among the most common neurodegenerative diseases in which a role for immune cells in its etiology is only beginning to be realized. AD has two molecular pathological hallmarks: the presence of amyloid beta (Aβ) plaques and tau tangles within neuronal cells. The amyloid plaques can be modeled in mice by inserting transgenes that enhance human Aβ to create plaques inside the mouse brain. Amyloid precursor protein (Mo/HuAPP695swe) and a mutant human presenilin 1 (PS1-dE9) are two transgenes inserted into C57BL/6 mice that results in Aβ plaques in the hippocampus and cortex. This mouse is known as the APP/PS1 mouse. Immune cell infiltration into the brain is associated with aged APP/PS1 mice. However, comprehensive immune profiling has not been performed in this model to outline immune cell effectors potentially involved in neuropathology. Therefore, we employed spectral flow cytometry to investigate the immune cell profile in 24-month-old APP/PS1 mice compared to an age matched control. There are increased numbers of CD8 T, γδ T, and NK cells expressing perforin and other activation related surface markers. Using small animal MRI imaging, we also observed enlarged ventricles mice indicative of brain atrophy in APP/PS1 mice when compared to wildtype age matched control. These results provide insight into the regional immune cell infiltration of the brain, setting the stage for further analysis of immune mechanisms of neuropathology in experimental AD.

The fluorescence mKate2 labeling as a visualizing system for monitoring small extracellular vesicles.

Small extracellular vesicles (sEVs) are nanosized vesicles enclosed in a lipid membrane released by nearly all cell types. sEVs have been considered as reliable biomarkers for diagnostics and effective carriers. Despite the clear importance of sEV functionality, sEV research faces challenges imposed by the small size and precise imaging of sEVs. Recent advances in live and high-resolution microscopy, combined with efficient labeling strategies, enable us to investigate the composition and behavior of EVs within living organisms. Here, a modified sEVs was generated with a near infrared fluorescence protein mKate2 using a VSVG viral pseudotyping-based approach for monitoring sEVs. An observed was made that the mKate2-tagged protein can be incorporated into the membranes of sEVs without altering their physical properties. In vivo imaging demonstrates that sEVs labeled with mKate2 exhibit excellent brightness and high photostability, allowing the acquisition of long-term investigation comparable to those achieved with mCherry labeling. Importantly, the mKate2-tagged sEVs show a low toxicity and exhibit a favorable safety profile. Furthermore, the co-expression of mKate2 and rabies virus glycoprotein (RVG) peptide on sEVs enables brain-targeted visualization, suggesting the mKate2 tag does not alter the biodistribution of sEVs. Together, the study presents the mKate2 tag as an efficient tracker for sEVs to monitor tissue-targeting and biodistribution in vivo.

Machine Teaching Allows for Rapid Development of Automated Systems for Retinal Lesion Detection From Small Image Datasets.

Machine teaching, a machine learning subfield, may allow for rapid development of artificial intelligence systems able to automatically identify emerging ocular biomarkers from small imaging datasets. We sought to use machine teaching to automatically identify retinal ischemic perivascular lesions (RIPLs) and subretinal drusenoid deposits (SDDs), two emerging ocular biomarkers of cardiovascular disease. IRB approval was obtained. Four small datasets of SD-OCT B-scans were used to train and test two distinct automated systems, one identifying RIPLs and the other identifying SDDs. An open-source interactive machine-learning software program, RootPainter, was used to perform annotation and training simultaneously over a 6-hour period. For SDDs at the B-scan level, test-set accuracy = 92%, sensitivity = 100%, specificity = 88%, positive predictive value (PPV) = 82%, and negative predictive value (NPV) = 100%. For RIPLs at the B-scan level, test-set accuracy = 90%, sensitivity = 60%, specificity = 93%, PPV = 50%, and NPV = 95%. Machine teaching demonstrates promise within ophthalmic imaging to rapidly allow for automated identification of novel biomarkers from small image datasets. [Ophthalmic Surg Lasers Imaging Retina 2024;55:XX-XX.].

Automatic generation method for long-focal-length unobscured freeform optical systems with small volume

The existing design methods for long-focal-length unobscured freeform systems rarely consider the imaging quality requirements and volume constraints simultaneously, causing most of the final designs to not fulfill the requirement of light weight. This study proposes a method to automatically design a long-focal-length unobscured reflective system that satisfies volume constraints while maintaining high imaging quality. First, a method to adaptively set the structural parameter range is proposed, and multiple parameters for different systemic specifications can be effectively calculated within it. Subsequently, the systemic volume and area functions are constructed using the ray tracing method, where the tilt angles, distances between mirrors, and radii of curvature of the mirrors are chosen as the optimization parameters. Third, a comprehensive objective function is jointly established combining ray obscuration and convergence as performance evaluation factors. Then, the structural parameters of a long-focal-length unobscured system with small volume are easily obtained via the simulated annealing method. Finally, the improved W-W method is used to further enhance the imaging quality of the system, and an unobscured freeform reflective optical system with three mirrors is automatically generated. Experimental results demonstrate that our method can automatically calculate the parameter ranges to facilitate the search for structural parameters, and effectively design the long-focal-length unobscured freeform systems with small volume and high imaging quality.

Small animal PET imaging with the 68Ga-labeled pH (low) insertion peptide-like peptide YJL-4 in a triple-negative breast cancer mouse model

BackgroundThe aim of this study was to prepare a novel 68Ga-labeled pH (low) insertion peptide (pHLIP)-like peptide, YJL-4, and determine its value for the early diagnosis of triple-negative breast cancer (TNBC) via in vivo imaging of tumor-bearing nude mice. The novel peptide YJL-4 was designed using a template-assisted method and synthesized by solid-phase peptide synthesis. After modification with the chelator 1,4,7‑triazacyclononane-N,N′,N″-triacetic acid (NOTA), the peptide was labeled with 68Ga. Then, the biodistribution of 68Ga-YJL-4 in tumor-bearing nude mice was investigated, and the mice were imaged by small animal positron emission tomography (PET).ResultsThe radiochemical yield and radiochemical purity of 68Ga-YJL-4 were 89.5 ± 0.16% and 97.95 ± 0.06%, respectively. The biodistribution of 68Ga-YJL-4 in tumors (5.94 ± 1.27% ID/g, 6.72 ± 1.69% ID/g and 4.54 ± 0.58% ID/g at 1, 2 and 4 h after injection, respectively) was significantly greater than that of the control peptide in tumors at the corresponding time points (P < 0.01). Of the measured off-target organs, 68Ga-YJL-4 was highly distributed in the liver and blood. The small animal PET imaging results were consistent with the biodistribution results. The tumors were visualized by PET at 2 and 4 h after the injection of 68Ga-YJL-4. No tumors were observed in the control group.ConclusionsThe novel pHLIP family peptide YJL-4 can adopt an α-helical structure for easy insertion into the cell membrane in an acidic environment. 68Ga-YJL-4 was produced in high radiochemical yield with good stability and can target TNBC tissue. Moreover, the strong concentration of radioactive 68Ga-YJL-4 in the abdomen does not hinder the imaging of early TNBC.

Sub-pixel target fine spatial feature extraction method based on aperture coding and micro-scanning imaging mechanism

The small imaging size of targets over long distances results in the loss of geometry and spatial features. Current methods are subject to sampling limitations and cannot accurately capture the spatial features of sub-pixel targets. This paper proposes a method to accurately locate and extract the fine spatial features of sub-pixel targets through aperture coding and micro-scanning imaging. First, the formation mechanism of imaging features for sub-pixel targets is analyzed. Second, the optical aperture is anisotropically coded in different directions to modulate the spreading spots of the target. The primary spreading direction and the center of the anisotropic spreading spots are extracted. The contour and the location of the target are determined from the spreading length and the intersections of the primary spreading directions. Then, the target is sampled by different detector units through various micro-scanning offsets. The pixel units containing different sub-pixel components of the target after offset are determined based on the location results. The fine spatial distribution of the sub-pixel target is reconstructed based on the intensity variations in the pixel units containing the target. Finally, the accuracy of the sub-pixel target fine spatial feature extraction method is validated. The results show a sub-pixel localization error of less than 0.02 and an effective improvement of the sub-pixel target spatial resolution. This paper provides significant potential for improving the ability to capture spatial features of targets over long distances.

Enhancing performance of vision transformers on small datasets through local inductive bias incorporation

Vision transformers (ViTs) achieve remarkable performance on large datasets, but tend to perform worse than convolutional neural networks (CNNs) when trained from scratch on smaller datasets, possibly due to a lack of local inductive bias in the architecture. Recent studies have therefore added locality to the architecture and demonstrated that it can help ViTs achieve performance comparable to CNNs in the small-size dataset regime. Existing methods, however, are architecture-specific or have higher computational and memory costs. Thus, we propose a module called Local InFormation Enhancer (LIFE) that extracts patch-level local information and incorporates it into the embeddings used in the self-attention block of ViTs. Our proposed module is memory and computation efficient, as well as flexible enough to process auxiliary tokens such as the classification and distillation tokens. Empirical results show that the addition of the LIFE module improves the performance of ViTs on small image classification datasets. We further demonstrate how the effect can be extended to downstream tasks, such as object detection and semantic segmentation. In addition, we introduce a new visualization method, Dense Attention Roll-Out, specifically designed for dense prediction tasks, allowing the generation of class-specific attention maps utilizing the attention maps of all tokens. The code for this project is available on Github (https://github.com/NeurAI-Lab/LIFEhttps://github.com/NeurAI-Lab/LIFE).

Global contextual representation via graph-transformer fusion for hepatocellular carcinoma prognosis in whole-slide images

Current methods of digital pathological images typically employ small image patches to learn local representative features to overcome the issues of computationally heavy and memory limitations. However, the global contextual features are not fully considered in whole-slide images (WSIs). Here, we designed a hybrid model that utilizes Graph Neural Network (GNN) module and Transformer module for the representation of global contextual features, called TransGNN. GNN module built a WSI-Graph for the foreground area of a WSI for explicitly capturing structural features, and the Transformer module through the self-attention mechanism implicitly learned the global context information. The prognostic markers of hepatocellular carcinoma (HCC) prognostic biomarkers were used to illustrate the importance of global contextual information in cancer histopathological analysis. Our model was validated using 362 WSIs from 355 HCC patients diagnosed from The Cancer Genome Atlas (TCGA). It showed impressive performance with a Concordance Index (C-Index) of 0.7308 (95% Confidence Interval (CI): (0.6283–0.8333)) for overall survival prediction and achieved the best performance among all models. Additionally, our model achieved an area under curve of 0.7904, 0.8087, and 0.8004 for 1-year, 3-year, and 5-year survival predictions, respectively. We further verified the superior performance of our model in HCC risk stratification and its clinical value through Kaplan–Meier curve and univariate and multivariate COX regression analysis. Our research demonstrated that TransGNN effectively utilized the context information of WSIs and contributed to the clinical prognostic evaluation of HCC.

Human figure detection in Han portrait stone images via enhanced YOLO-v5

The unearthed Han Dynasty portrait stones are an important part of China’s ancient artistic heritage, and detecting human images in these stones is a critical prerequisite for studying their artistic value. However, high-precision target detection techniques often result in a large number of parameters, making them unsuitable for portable devices. In this work, we propose a new human image target detection model based on an enhanced YOLO-v5. We discovered that the complex backgrounds, dense group targets, and significant scale variations of targets within large scenes in portrait stones present significant challenges for human target image detection. Therefore, we first incorporated the SPD-Conv convolution and Coordinate Attention self-attention mechanism modules into the YOLO-v5 architecture, aiming to enhance the model’s recognition precision for small target images within Han portrait stones and strengthen its resistance to background disturbances. Moreover, we introduce DIoU NMS and Alpha-IoU Loss to improve the detector’s performance in dense target scenarios, reducing the omission of densely packed objects. Finally, the experimental results from our collected dataset of Han Dynasty stone figure images demonstrate that our method achieves fast convergence and high recognition accuracy. This approach can be better applied to the target detection tasks of special character images in complex backgrounds.

Automated detection of small bowel lesions based on capsule endoscopy using deep learning algorithm

BackgroundIn order to overcome the challenges of lesion detection in capsule endoscopy (CE), we improved the YOLOv5-based deep learning algorithm and established the CE-YOLOv5 algorithm to identify small bowel lesions captured by CE. MethodsA total of 124,678 typical abnormal images from 1,452 patients were enrolled to train the CE-YOLOv5 model. Then 298 patients with suspected small bowel lesions detected by CE were prospectively enrolled in the testing phase of the study. Small bowel images and videos from the above 298 patients were interpreted by the experts, non-experts and CE-YOLOv5, respectively. ResultsThe sensitivity of CE-YOLOv5 in diagnosing vascular lesions, ulcerated/erosive lesions, protruding lesions, parasite, diverticulum, active bleeding and villous lesions based on CE videos was 91.9 %, 92.2 %, 91.4 %, 93.1 %, 93.3 %, 95.1 %, and 100 % respectively. Furthermore, CE-YOLOv5 achieved specificity and accuracy of more than 90 % for all lesions. Compared with experts, the CE-YOLOv5 showed comparable overall sensitivity, specificity and accuracy (all P > 0.05). Compared with non-experts, the CE-YOLOv5 showed significantly higher overall sensitivity (P < 0.0001) and overall accuracy (P < 0.0001), and a moderately higher overall specificity (P = 0.0351). Furthermore, the time for AI-reading (5.62 ± 2.81 min) was significantly shorter than that for the other two groups (both P < 0.0001). ConclusionsCE-YOLOv5 diagnosed small bowel lesions in CE videos with high sensitivity, specificity and accuracy, providing a reliable approach for automated lesion detection in real-world clinical practice.