Image Slices Research Articles

Self-supervised learning for chest computed tomography: training strategies and effect on downstream applications.

Self-supervised pre-training can reduce the amount of labeled training data needed by pre-learning fundamental visual characteristics of the medical imaging data. We investigate several self-supervised training strategies for chest computed tomography exams and their effects on downstream applications. We benchmark five well-known self-supervision strategies (masked image region prediction, next slice prediction, rotation prediction, flip prediction, and denoising) on 15M chest computed tomography (CT) slices collected from four sites of the Mayo Clinic enterprise, United States. These models were evaluated for two downstream tasks on public datasets: pulmonary embolism (PE) detection (classification) and lung nodule segmentation. Image embeddings generated by these models were also evaluated for prediction of patient age, race, and gender to study inherent biases in models' understanding of chest CT exams. The use of pre-training weights especially masked region prediction-based weights, improved performance, and reduced computational effort needed for downstream tasks compared with task-specific state-of-the-art (SOTA) models. Performance improvement for PE detection was observed for training dataset sizes as large as with a maximum gain of 5% over SOTA. The segmentation model initialized with pre-training weights learned twice as fast as the randomly initialized model. While gender and age predictors built using self-supervised training weights showed no performance improvement over randomly initialized predictors, the race predictor experienced a 10% performance boost when using self-supervised training weights. We released self-supervised models and weights under an open-source academic license. These models can then be fine-tuned with limited task-specific annotated data for a variety of downstream imaging tasks, thus accelerating research in biomedical imaging informatics.

Comparing assisting technologies for proficiency in cardiac morphology: 3D printing and mixed reality versus CT slice images for morphological understanding of congenital heart defects by medical students.

Learning cardiac morphology largely involves spatial abilities and studies indicate benefits from innovative 3D visualization technologies that speed up and increase the learning output. Studies comparing these teaching tools and their educational output are rare and few studies include complex congenital heart defects. This study compared the effects of 3D prints, mixed reality (MR) viewing of 3D meshes and standard cardiac CT slice images on medical students' understanding of complex congenital heart defect morphology, measuring both objective level of understanding and subjective educational experience. The objective of this study was to compare morphological understanding and user experiences of 3D printed models, MR 3D visualization and axial 2D CT slices, in medical students examining morphological details in complex congenital heart defects. Medical students in the median 4th year of study (range 2nd to 6th) examined three of five different complex congenital heart defects by three different modalities: 3D printed model, MR viewed 3D mesh, and cardiac CT slices, answering a questionnaire on morphology and user experience. Time to complete task, diagnostic accuracy, and user experience data were collected and compared on group level. Task times were similar for all modalities. The percentage of correct answers was higher with MR visualization, which was also the preferred modality overall. Medical students both prefer and better understand the morphology of complex congenital heart disease with 3D models viewed using MR, without spending more time than with 3D prints or 2D CT images.

CT image segmentation of foxtail millet seeds based on semantic segmentation model VGG16-UNet

Foxtail millet is an important minor cereal crop rich in nutrients. Due to the small size of its seeds, there is little information on the diversity of its seed structure among germplasms, limiting the identification of genes controlling seed development and germination. This paper utilized X-ray computed tomography (CT) scanning technology and deep learning models to reveal the microstructure of foxtail millet seeds, gaining insights into their internal features, distribution, and composition. A total of 100 foxtail millet varieties were scanned with X-ray computed tomography to obtain 3D reconstruction images and slices. Pre-processing steps were adopted to improve image segmentation accuracy, including noise reduction, rotation, contrast enhancement, and brightness enhancement. The experiment revealed that traditional OpenCV image processing methods failed to achieve precise segmentation, whereas deep learning models exhibited outstanding performance in segmenting seed CT slice images. We compared UNet, PSPNet, and DeepLabV3 models, selected different backbones and optimizers based on the dataset, and continuously adjusted learning rates and maximum training epochs to train the models. Results demonstrated that VGG16-UNet achieved an accuracy of 99.19% on the foxtail millet seed CT slice image dataset, outperforming PSPNet and DeepLabV3 models. Compared to ResNet-UNet, VGG16-UNet shows an improvement of approximately 3.18% in accuracy, demonstrating superior performance in accurately segmenting the inner glume, outer glume, embryo, and endosperm under various adhesion conditions. Accurate segmentation of foxtail millet CT images enables analysis of embryo size, endosperm size, and glume thickness, which impact germination, growth, and nutrition. This study fills a gap in small grain structure research, offering insights to optimize agriculture and molecular breeding for improved yield and quality.

CSSANet: A Channel Shuffle Slice-Aware Network for Pulmonary Nodule Detection

Lung cancer stands as the leading cause of cancer related mortality worldwide. Precise and automated identification of lung nodules through 3D Computed Tomography (CT) scans is an essential part of screening for lung cancer effectively. Due to the small size of pulmonary nodules and the close correlation between neighboring slices of 3D CT images, most of the existing methods only consider the characteristics of a single slice, thus easily lead to insufficient detection accuracy of pulmonary nodules. To solve this problem, this paper proposes a Channel Shuffle Slice-Aware Network (CSSANet), which aims to fully exploit the spatial correlation between slices and effectively utilize the intra-slice features and inter-slice contextual information to achieve accurate detection of lung nodules. Specifically, we design a Group Shuffle Attention module (GSA module) to fuse the inter-slice feature in order to enhance the discrimination and extraction of corresponding shape information of distinct nodules in the same group of slices. Experiments and ablation study on a publicly available LUNA16 dataset demonstrate that the proposed method can enhance the detection sensitivity effectively. The Competition Performance Metric (CPM) score of 89.8% is superior over other representative detection models.

Optimal transport assisted full waveform inversion for multiparameter imaging of soft tissues in ultrasound computed tomography

Ultrasound computed tomography (USCT) has emerged as a promising platform for imaging tissue properties, offering non-ionizing and operator-independent capabilities. In this work, we demonstrate the feasibility of obtaining quantitative images of multiple acoustic parameters (sound speed and impedance) for soft tissues using full waveform inversion (FWI), which are justified with both numerical and experimental cases. A 3D reconstruction based on a series of 2D slice images is presented for the experimental case of ex vivo soft tissues. To improve the robustness of the reconstruction process, a hierarchical FWI strategy is adopted, gradually iterating from low to high frequencies. In parallel, we employ a graph-space optimal transport misfit function, avoiding convergence into local minima and minimizing inversion artifacts caused by skin-related supercritical reflections. Our method first carries out sound speed inversion based on transmitted waves in the low and middle frequency bands, and then uses all types of waves in the high frequency band for simultaneous inversion of both sound speed and impedance. Compared to conventional strategies, the proposed approach can accurately reconstruct physical models consistent with the actual soft tissue sample. These high-resolution ultrasound images of acoustic parameters are promising to allow for quantitative differentiation among different types of tissues (e.g., muscles and fats). These results have significant implications for advancing our understanding of tissue properties and for potentially contributing to disease diagnosis through USCT, which is a flexible and cost-effective alternative to X-ray computed tomography or magnetic resonance imaging at no significant sacrifices for resolution.

Establishing a Whole-Cell Configuration for Two-Photon Calcium Imaging of Brain Slices

Establishing a Whole-Cell Configuration for Two-Photon Calcium Imaging of Brain Slices

Establishing a Whole-Cell Configuration for Two-Photon Calcium Imaging of Brain Slices

Establishing a Whole-Cell Configuration for Two-Photon Calcium Imaging of Brain Slices

Ultra-High-Resolution Photon-Counting Detector CT Benefits Visualization of Abdominal Arteries: A Comparison to Standard-Reconstruction.

This study aimedto investigate the potential benefit of ultra-high-resolution (UHR) photon-counting detector CT (PCD-CT) angiography in visualization of abdominal arteries in comparison to standard-reconstruction (SR) images of virtual monoenergetic images (VMI) at low kiloelectron volt (keV).We prospectively included 47 and 47 participants to undergo contrast-enhanced abdominal CT scans within UHR mode on a PCD-CT systemusing full-dose (FD) and low-dose (LD) protocols, respectively. The data were reconstructed into six series of images: FD_UHR_Br48, FD_UHR_Bv56, FD_UHR_Bv60, FD_SR_Bv40, LD_UHR_Bv48, and LD_SR_Bv40. The UHR reconstructions were performed with three kernels (Bv48, Bv56, and Bv60) within 0.2mm. The SR were virtual monoenergetic imaging reconstruction with Bv40 kernel at 40-keV within 1mm. Each series of axial images were reconstructed into coronal and volume-rendered images. The signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR) of seven arteries were measured. Three radiologists assessed the image quality, and visibility of nine arteries on all the images.SNR and CNR values of SR images were significantly higher than those of UHR images (P < 0.001). The SR images have higher ratings in image noise (P < 0.001), but the FD_UHR_Bv56 and FD_UHR_Bv60 images has higher rating in vessel sharpness (P < 0.001). The overall quality was not significantly different among FD_VMI_40keV, LD_VMI_40keV, FD_UHR_Bv48, and LD_UHR_Bv48 images (P > 0.05) but higher than those of FD_UHR_Bv56 and FD_UHR_Bv60 images (P < 0.001). There is no significant difference of nine abdominal arteries among six series of images of axial, coronal and volume-rendered images (P > 0.05).To conclude, 1-mm SR image of VMI at 40-keV is superior to 0.2-mm UHR regardless of which kernel is used to visualize abdominal arteries, while 0.2-mm UHR image using a relatively smooth kernel may allow similar image quality and artery visibility when thinner slice image is warranted.

A Deep Learning-based Pipeline for Segmenting the Cerebral Cortex Laminar Structure in Histology Images.

Characterizing the anatomical structure and connectivity between cortical regions is a critical step towards understanding the information processing properties of the brain and will help provide insight into the nature of neurological disorders. A key feature of the mammalian cerebral cortex is its laminar structure. Identifying these layers in neuroimaging data is important for understanding their global structure and to help understand the connectivity patterns of neurons in the brain. We studied Nissl-stained and myelin-stained slice images of the brain of the common marmoset (Callithrix jacchus), which is a new world monkey that is becoming increasingly popular in the neuroscience community as an object of study. We present a novel computational framework that first acquired the cortical labels using AI-based tools followed by a trained deep learning model to segment cerebral cortical layers. We obtained a Euclidean distance of for the cortical labels acquisition, which was in the acceptable range by computing the half Euclidean distance of the average cortex thickness ( ). We compared our cortical layer segmentation pipeline with the pipeline proposed by Wagstyl et al. (PLoS biology, 18(4), e3000678 2020) adapted to 2D data. We obtained a better mean percentile Hausdorff distance (95HD) of . Whereas a mean 95HD of was obtained from Wagstyl et al. We also compared our pipeline's performance against theirs using their dataset (the BigBrain dataset). The results also showed better segmentation quality, Jaccard Index acquired from our pipeline, while was stated in their paper.

Accuracy of pelvic bone segmentation for 3d printing: a study of segmentation accuracy based on anatomic landmarks to evaluate the influence of the observer.

3D printing has a wide range of applications and has brought significant change to many medical fields. However, ensuring quality assurance (QA) is essential for patient safety and requires a QA program that encompasses the entire production process. This process begins with imaging and continues on with segmentation, which is the conversion of Digital Imaging and Communications in Medicine (DICOM) data into virtual 3D-models. Since segmentation is highly influenced by manual intervention the influence of the users background on segmentation accuracy should be thoroughly investigated. Seventeen computed tomography (CT) scans of the pelvis with physiological bony structures were identified, anonymized, exported as DICOM data sets, and pelvic bones were segmented by four observers with different backgrounds. Landmarks were measured on DICOM images and in the segmentations. Intraclass correlation coefficients (ICCs) were calculated to assess inter-observer agreement, and the trueness of the segmentation results was analyzed by comparing the DICOM landmark measurements with the measurements of the segmentation results. The correlation between segmentation trueness and segmentation time was analyzed. The lower limits of the 95% confidence intervals of the ICCs for the seven landmarks analyzed ranged from 0.511 to 0.986. The distance between the iliac crests showed the highest agreement between observers, while the distance between the ischial tuberosities showed the lowest. The distance between the upper edge of the symphysis and the promontory showed the lowest deviation between DICOM measurements and segmentation measurements (mean deviations < 1mm), while the intertuberous distance showed the highest deviation (mean deviations 14.5-18.2mm). Investigators with diverse backgrounds in segmentation and varying experience with slice images achieved pelvic bone segmentations with landmark measurements of mostly high agreement in a setup with high realism. In contrast, high variability was observed in the segmentation of the coccyx. In general, interobserver agreement was high, but due to measurement inaccuracies, landmark-based approaches cannot conclusively show that segmentation accuracy is within a clinically tolerable range of 2mm for the pelvis. If the segmentation is performed by a very inexperienced user, the result should be reviewed critically by the clinician in charge.

Analysing the informative content of the upper jaw pathology visual examination

Objective. To conduct a comparative analysis of the informative content of cone beam computed tomography (CBCT), multispiral computed tomography (MSCT) and paranasal sinuses half-axial X-ray images in the upper jaw pathology diagnosis. Materials and methods. 96 CBCT and MSCT scans of the upper jaw and the similar number of X-ray images of the same patients were analyzed. Results. The analysis of a three-dimensional image and native slice reading in the Planmeca programme makes it possible to determine metric parameters, localization, prevalence of the pathological process and involvement of the upper jaw structures validly. Software features of the Radiant and OneDemand3D had error interval of р ≥ 0,2 mm while analyzing the same criteria. Half-axial X-ray examination demonstrated the lowest informative content. Conclusion. The most informative method of visualization of the upper jaw connective tissue structures, pathological processes, their borders and mutual involvement is CBCT. This is predetermined by the interface functionalities.

Histological Validation of 3D Variable Flip Angle TSE Multi-Contrast Magnetic Resonance Vessel Wall Imaging in Characterizing Carotid Vulnerable Atherosclerotic Plaques

BackgroundAccurate assessment of the vulnerability of carotid atherosclerotic plaques is crucial for stroke prevention. The three-dimensional (3D) magnetic resonance (MR) vessel wall imaging (VWI) has been increasingly employed to evaluate carotid plaques due to its extensive coverage and isotropic high spatial resolution. However, the accuracy of such technique lacks validation by histology. ObjectiveThis study aims to validate the accuracy of 3D multi-contrast MR VWI used variable-flip-angle (VFA) and turbo spin echo (TSE) readout in identifying vulnerable carotid plaques, using histological analysis as a reference. MethodsTwenty-one male patients (mean age: 64.4 ± 7.2 years) scheduled for carotid endarterectomy (CEA) were recruited for this study. All patients underwent carotid multi-contrast MR VWI, including 3D T1- and T2-weighted variable flip angle-based turbo spin echo (VFA-TSE) sequences, as well as 3D time of flight (TOF) MR angiography (MRA), using a 3.0T MR system. Histological processing was performed for carotid plaque specimens. The presence or absence, along with the area measurements, of lipid-rich necrotic core (LRNC), intraplaque hemorrhage (IPH), and calcifications (CA) were independently evaluated on both MR images and histological sections. Cohen’s kappa (κ) analysis was utilized to determine the agreement between 3D multi-contrast MR VWI and histology in identifying carotid plaque compositions before and after excluding compositions bellow certain size threshold. Spearman’s correlation analysis was also conducted to assess the agreement in quantifying plaque compositions. ResultsA total of 81 slices of MR images were successfully matched with histological sections. Moderate to almost perfect agreements were observed between 3D MR VWI and histology in the identification of LRNC (κ: 0.85 and 0.89), IPH (κ: 0.65 and 0.69), and CA (κ: 0.46 and 0.62) before and after excluding compositions smaller than 0.79 mm2. Strong to very strong correlations were found in the quantification of plaque compositions including LRNC (r=0.88), IPH (r=0.80), and CA (r=0.74) between MR imaging and histology. ConclusionThe 3D VFA-TSE multi-contrast MR VWI is capable of accurately characterizing vulnerable carotid atherosclerotic plaques.

Moving ships detection via the trajectory feature extraction from spatiotemporal slices of infrared maritime videos

In practical application scenarios, maritime infrared videosoften contain different types of sea state scenes, weather conditions, shooting time, shooting distance, etc. In these different types of maritime videos, ship targets have great differences in size, grayscaledistribution and contrast, which brings difficulties to ship target detection.At the same time, the diversity of fluctuation, grayscale distribution and reflected light of the sea surface will bring unpredictable interference noise to ship target detection.How to accurately detect ship targets in complex and changeable maritime infrared videos is a challenging task and research focus.The key to achieving accurate ship detection is to extract robust target features that can effectively distinguish targets from all thebackground noises.In this paper, a novel infrared video ship target detection algorithm based on spatiotemporal slice target trajectory features extraction is proposed. The algorithm is very sensitive to real targets and has excellentanti-noise ability.The main innovation of the algorithm is to extract the target trajectory feature from the spatiotemporal slice of the sequence image and generate the trajectory feature map.We use the trajectory texture formed by the ship target in the spatiotemporal slice to extract the target feature, which can greatly suppress the background noise.The adaptive dilation linear model algorithm can effectively detect the target trajectory line in the spatiotemporal slice. In addition, we also make full use of the gradient of the target trajectory line to distinguish different target trajectory pixels, and propose an adaptive iterative dilation target region localization algorithm combined with gradient consistency.For object segmentation, we calculate the segmentation double-threshold using the adjacent surrounding background pixels of the target, so as to achievetarget segmentationof multiple grayscaledistribution types. Finally, in the comparison experiment, our algorithm shows superior target detection performance, especially when detecting ships from a large number of highlighted sea clutter background, the robust anti-noise ability of the algorithm can be highlighted.

Multiple myeloma segmentation net (MMNet): an encoder-decoder-based deep multiscale feature fusion model for multiple myeloma segmentation in magnetic resonance imaging.

Patients with multiple myeloma (MM), a malignant disease involving bone marrow plasma cells, shows significant susceptibility to bone degradation, impairing normal hematopoietic function. The accurate and effective segmentation of MM lesion areas is crucial for the early detection and diagnosis of myeloma. However, the presence of complex shape variations, boundary ambiguities, and multiscale lesion areas, ranging from punctate lesions to extensive bone damage, presents a formidable challenge in achieving precise segmentation. This study thus aimed to develop a more accurate and robust segmentation method for MM lesions by extracting rich multiscale features. In this paper, we propose a novel, multiscale feature fusion encoding-decoding model architecture specifically designed for MM segmentation. In the encoding stage, our proposed multiscale feature extraction module, dilated dense connected net (DCNet), is employed to systematically extract multiscale features, thereby augmenting the model's sensing field. In the decoding stage, we propose the CBAM-atrous spatial pyramid pooling (CASPP) module to enhance the extraction of multiscale features, enabling the model to dynamically prioritize both channel and spatial information. Subsequently, these features are concatenated with the final output feature map to optimize segmentation outcomes. At the feature fusion bottleneck layer, we incorporate the dynamic feature fusion (DyCat) module into the skip connection to dynamically adjust feature extraction parameters and fusion processes. We assessed the efficacy of our approach using a proprietary dataset of MM, yielding notable advancements. Our dataset comprised 753 magnetic resonance imaging (MRI) two-dimensional (2D) slice images of the spinal regions from 45 patients with MM, along with their corresponding ground truth labels. These images were primarily obtained from three sequences: T1-weighted imaging (T1WI), T2-weighted imaging (T2WI), and short tau inversion recovery (STIR). Using image augmentation techniques, we expanded the dataset to 3,000 images, which were employed for both model training and prediction. Among these, 2,400 images were allocated for training purposes, while 600 images were reserved for validation and testing. Our method showed increase in the intersection over union (IoU) and Dice coefficients by 7.9 and 6.7 percentage points, respectively, as compared to the baseline model. Furthermore, we performed comparisons with alternative image segmentation methodologies, which confirmed the sophistication and efficacy of our proposed model. Our proposed multiple myeloma segmentation net (MMNet), can effectively extract multiscale features from images and enhance the correlation between channel and spatial information. Furthermore, a systematic evaluation of the proposed network architecture was conducted on a self-constructed, limited dataset. This endeavor holds promise for offering valuable insights into the development of algorithms for future clinical applications.

Automatic detection of hidden defects and qualification of additively manufactured parts using X-ray computed tomography and computer vision

Additive manufacturing (AM) has the unique capability to produce parts with complex three-dimensional structures with features that are internal to the part and hidden from view. Such internal features are difficult to inspect and may have defects that affect the function of a part. X-ray computed tomography (CT) is one of the only methods for nondestructive inspection (NDI) of the interior of AM parts. This paper explores how machine learning (ML) methods can be used to analyze CT scans of AM parts to automatically identify the presence of defects in geometric features. We designed a nozzle part with internal three-dimensional (3D) channels and introduced five different synthetic defects whose presence could affect the nozzle performance. A resin-based AM process fabricated 155 nozzle parts that included both defect-free parts and parts with synthetic defects. CT scans were collected for each part and processed into 68,510 image cross sections. The extracted images were used to train an ML model based on the ResNet34 architecture. The model can automatically identify and classify defects in individual CT slice images with over 98% accuracy. High model accuracy is possible with training on as few as 30 parts. The research demonstrates the potential of ML methods to automatically identify hidden defects and qualify AM parts using X-ray CT scans.

Ex Vivo Culture and Imaging of Oculomotor Slices from Transgenic Mouse Embryos

Ex Vivo Culture and Imaging of Oculomotor Slices from Transgenic Mouse Embryos

Ex Vivo Culture and Imaging of Oculomotor Slices from Transgenic Mouse Embryos

Ex Vivo Culture and Imaging of Oculomotor Slices from Transgenic Mouse Embryos

Removal of ring artifacts in neutron tomography image slices

Removal of ring artifacts in neutron tomography image slices

Structural Effects on the Energy Disposal and Atomic Photofragment Alignment for the Photodissociation of Alkyl Iodides at Excitation Wavelengths of 254 and 268 nm.

This work represents a step forward in the understanding of the widely studied photodynamics of alkyl iodides in the first absorption band. Ultraviolet (UV) photodissociation of several alkyl iodides (RI), specifically, a series of linear and ramified molecules with R = CnH2n+1, n = 1-4, at excitation wavelengths of 254 and 268 nm, which correspond to the maximum of the first absorption A-band, has been studied by combining resonance-enhanced multiphoton ionization (REMPI) detection of atomic photofragments I(2P3/2) and I*(2P1/2) and of pulsed slice imaging. Detailed examination of the total translational energy distributions of both atomic photofragments has been combined with stereodynamical information on the process obtained from the anisotropy β and alignment a02(∥) and Re[a12(∥, ⊥)] parameters to provide a description of the role played by the molecular structure of alkyl iodides in adiabatic and, especially, in nonadiabatic photodissociation dynamics through conical intersections or avoided crossings. The present results suggest that the linear structures couple more efficiently with the pure C-I reaction coordinate, whereas for the branched structures, the coupling with additional vibrational (bending) modes gains importance, showing the dissociation process a multidimensional character. In addition, a large degree of cofragment rotational alignment has been found for the small linear CH3I and C2H5I and, unexpectedly, for the branched t-C4H9I (C3v symmetry), whereas the rest of the alkyl iodides show low alignment parameters.

Non-contrasted computed tomography (NCCT) based chronic thromboembolic pulmonary hypertension (CTEPH) automatic diagnosis using cascaded network with multiple instance learning

Objective.The diagnosis of chronic thromboembolic pulmonary hypertension (CTEPH) is challenging due to nonspecific early symptoms, complex diagnostic processes, and small lesion sizes. This study aims to develop an automatic diagnosis method for CTEPH using non-contrasted computed tomography (NCCT) scans, enabling automated diagnosis without precise lesion annotation.Approach.A novel cascade network (CN) with multiple instance learning (CNMIL) framework was developed to improve the diagnosis of CTEPH. This method uses a CN architecture combining two Resnet-18 CNN networks to progressively distinguish between normal and CTEPH cases. Multiple instance learning (MIL) is employed to treat each 3D CT case as a 'bag' of image slices, using attention scoring to identify the most important slices. An attention module helps the model focus on diagnostically relevant regions within each slice. The dataset comprised NCCT scans from 300 subjects, including 117 males and 183 females, with an average age of 52.5 ± 20.9 years, consisting of 132 normal cases and 168 cases of lung diseases, including 88 cases of CTEPH. The CNMIL framework was evaluated using sensitivity, specificity, and the area under the curve (AUC) metrics, and compared with common 3D supervised classification networks and existing CTEPH automatic diagnosis networks.Main results. The CNMIL framework demonstrated high diagnostic performance, achieving an AUC of 0.807, accuracy of 0.833, sensitivity of 0.795, and specificity of 0.849 in distinguishing CTEPH cases. Ablation studies revealed that integrating MIL and the CN significantly enhanced performance, with the model achieving an AUC of 0.978 and perfect sensitivity (1.000) in normal classification. Comparisons with other 3D network architectures confirmed that the integrated model outperformed others, achieving the highest AUC of 0.8419.Significance. The CNMIL network requires no additional scans or annotations, relying solely on NCCT. This approach can improve timely and accurate CTEPH detection, resulting in better patient outcomes.