Journal of Medical Imaging最新文献

Purpose: We aim to provide a robust dataset for training automated systems to detect tuberculosis bacilli using Ziehl-Neelsen stained slides. By making this dataset available, a critical gap in the availability of public datasets that can be used for developing and testing artificial intelligence techniques for tuberculosis diagnosis is addressed. Our rationale is grounded in the urgent need for diagnostic tools that can enhance tuberculosis diagnosis quickly and efficiently, especially in resource-limited settings.

Approach: The Ziehl-Neelsen method was used to prepare 362 slides, which were manually read. According to the World Health Organization's guidelines for performing bacilloscopy for tuberculosis diagnosis, experts annotated each slide to diagnose it as negative or positive. In addition, selected images underwent a detailed annotation process aimed at pinpointing the location of each bacillus and cluster within each image.

Results: The database consists of three directories. The first contains all the images, separated by slide, and indicates whether it is negative or the number of crosses if positive, for each slide. The second directory contains the 502 images selected for training automated systems, with each bacillus's position annotated and the Python code used. All the image fragments (positive and negative patches) used in the models' training, validation, and testing stages are available in the third directory.

Conclusions: The development of this annotated image database represents a significant advancement in tuberculosis diagnosis. By providing a high-quality and accessible resource to the scientific community, we enhance existing diagnostic tools and facilitate the development of automated technologies.

Purpose: Segmentation of ultrasound images for medical diagnosis, monitoring, and research is crucial, and although existing methods perform well, they are limited by specific organs, tumors, and image devices. Applications of the Segment Anything Model (SAM), such as SAM-med2d, use a large number of medical datasets that contain only a small fraction of the ultrasound medical images.

Approach: In this work, we proposed a SAM-MedUS model for generic ultrasound image segmentation that utilizes the latest publicly available ultrasound image dataset to create a diverse dataset containing eight site categories for training and testing. We integrated ConvNext V2 and CM blocks in the encoder for better global context extraction. In addition, a boundary loss function is used to improve the segmentation of fuzzy boundaries and low-contrast ultrasound images.

Results: Experimental results show that SAM-MedUS outperforms recent methods on multiple ultrasound datasets. For the more easily datasets such as the adult kidney, it achieves 87.93% IoU and 93.58% dice, whereas for more complex ones such as the infant vein, IoU and dice reach 62.31% and 78.93%, respectively.

Conclusions: We collected and collated an ultrasound dataset of multiple different site types to achieve uniform segmentation of ultrasound images. In addition, the use of additional auxiliary branches ConvNext V2 and CM block enhances the ability of the model to extract global information and the use of boundary loss allows the model to exhibit robust performance and excellent generalization ability.

Purpose: The advancement of high-content optical microscopy has enabled the acquisition of very large three-dimensional (3D) image datasets. The analysis of these image volumes requires more computational resources than a biologist may have access to in typical desktop or laptop computers. This is especially true if machine learning tools are being used for image analysis. With the increased amount of data analysis and computational complexity, there is a need for a more accessible, easy-to-use, and efficient network-based 3D image processing system. The distributed and networked analysis of volumetric image data (DINAVID) system was developed to enable remote analysis of 3D microscopy images for biologists.

Approach: We present an overview of the DINAVID system and compare it to other tools currently available for microscopy image analysis. DINAVID is designed using open-source tools and has two main sub-systems, a computational system for 3D microscopy image processing and analysis and a 3D visualization system.

Results: DINAVID is a network-based system with a simple web interface that allows biologists to upload 3D volumes for analysis and visualization. DINAVID enables the image access model of a center hosting image volumes and remote users analyzing those volumes, without the need for remote users to manage any computational resources.

Conclusions: The DINAVID system, designed and developed using open-source tools, enables biologists to analyze and visualize 3D microscopy volumes remotely without the need to manage computational resources. DINAVID also provides several image analysis tools, including pre-processing and several segmentation models.

Purpose: Machine learning (ML) has been used to predict tumor progression post-stereotactic radiosurgery (SRS) based on pre-treatment MRI for brain metastasis (BM) patients, but there is variability in the definition of what constitutes progression. We aim to measure the magnitude of the change of performance of an ML model predicting post-SRS progression when various definitions of progression were used.

Approach: We collected pre- and post-SRS contrast-enhanced T1-weighted MRI scans from 62 BM patients ( $n=115$ BMs). We trained a random decision forest model using radiomic features extracted from pre-SRS scans to predict progression versus non-progression for each BM. We varied the definition of progression by changing (1) the follow-up period ( $<9$ , $<12$ , $<15$ , $<18$ , or $<24$ months); (2) the size change metric denoting progression ( $\ge 10\%$ , $\ge 15\%$ , $\ge 20\%$ , or $\ge 25\%$ in volume) or response assessment in neuro-oncology BM diameter ( $\ge 20\%$ ); and (3) whether BMs with treatment-related size changes (TRSCs) (pseudo-progression and/or radiation-necrosis) were labeled as progression. We measured performance using the area under the receiver operating characteristic curve (AUC).

Results: When we varied the follow-up period, size change metric, and TRSC labeling, the AUCs had ranges of 0.06 (0.69 to 0.75), 0.06 (0.69 to 0.75), and 0.08 (0.69 to 0.77), respectively. Radiomic feature importance remained similar.

Conclusions: Variability in the definition of BM progression has a measurable impact on the performance of an MRI radiomic-based ML model predicting post-SRS progression. A consistent, clinically relevant definition of post-SRS progression across studies would enable robust comparison of proposed ML systems, thereby accelerating progress in this field.

Purpose: We present a semi-supervised method for intestine segmentation to assist clinicians in diagnosing intestinal diseases. Accurate segmentation is essential for planning treatments for conditions such as intestinal obstruction. Although fully supervised learning performs well with abundant labeled data, the complexity of the intestine's spatial structure makes labeling time-intensive, resulting in limited labeled data. We propose a 3D segmentation network with a bidirectional teaching strategy to enhance segmentation accuracy using this limited dataset.

Method: The proposed semi-supervised method segments the intestine from computed tomography (CT) volumes using bidirectional teaching, where two backbones with different initial weights are trained simultaneously to generate pseudo-labels and employ unlabeled data, mitigating the challenge of limited labeled data. Intestine segmentation is further complicated by complex spatial features. To address this, we propose a lightweight multi-view symmetric network, which uses small-sized convolutional kernels instead of large ones to reduce parameters and capture multi-scale features from diverse perceptual fields, enhancing learning ability.

Results: We evaluated the proposed method with 59 CT volumes and repeated all experiments five times. Experimental results showed that the average Dice of the proposed method was 80.45%, the average precision was 84.12%, and the average recall was 78.84%.

Conclusions: The proposed method can effectively utilize large-scale unlabeled data with pseudo-labels, which is crucial in reducing the effect of limited labeled data in medical image segmentation. Furthermore, we assign different weights to the pseudo-labels to improve their reliability. From the result, we can see that the method produced competitive performance compared with previous methods.

Purpose: We aim to develop and validate a prediction model using a previously developed fully automated quantitative fissure integrity score (FIS) extracted from pre-treatment CT images to identify suitable candidates for endobronchial valve (EBV) treatment.

Approach: We retrospectively collected 96 anonymized pre- and post-treatment chest computed tomography (CT) exams from patients with moderate to severe emphysema and who underwent EBV treatment. We used a previously developed fully automated, deep learning-based approach to quantitatively assess the completeness of each fissure by obtaining the FIS for each fissure from each patient's pre-treatment CT exam. The response to EBV treatment was recorded as the amount of targeted lobe volume reduction (TLVR) compared with target lobe volume prior to treatment as assessed on the pre- and post-treatment CT scans. EBV placement was considered successful with a TLVR of $\ge 350\mathrm{cc}$ . The dataset was split into a training set ( $N=58$ ) and a test set ( $N=38$ ) to train and validate a logistic regression model using fivefold cross-validation; the extracted FIS of each patient's targeted treatment lobe was the primary CT predictor. Using the training set, a receiver operating characteristic (ROC) curve analysis and predictive values were quantified over a range of FIS thresholds to determine an optimal cutoff value that would distinguish complete and incomplete fissures, which was used to evaluate predictive values of the test set cases.

Results: ROC analysis of the training set provided an AUC of 0.83, and the determined FIS threshold was 89.5%. Using this threshold on the test set achieved an accuracy of 81.6%, specificity (Sp) of 90.9%, sensitivity (Sn) of 77.8%, positive predictive value (PPV) of 62.5%, and negative predictive value of 95.5%.

Conclusions: A model using the quantified FIS shows potential as a predictive biomarker for whether a targeted lobe will achieve successful volume reduction from EBV treatment.

Purpose: Accurate and efficient automatic segmentation of brain tumors is critical for diagnosis and treatment. However, the diversity in the appearance, location, and shape of brain tumors and their subregions, coupled with complex boundaries, presents significant challenges. We aim to improve segmentation accuracy by addressing limitations in V-Net, including insufficient utilization of multi-scale features and difficulties in managing complex spatial relationships and long-range dependencies.

Approach: We propose an improved network structure, dynamic convolution enhanced fusion axial V-Net (DCEF-AVNet), which integrates an enhanced feature fusion module and axial attention mechanisms. The feature fusion module integrates dynamic convolution with a redesigned skip connection strategy to effectively combine multi-scale features, reducing feature inconsistencies and improving representation capability. Axial attention mechanisms are introduced at encoder-decoder connections to manage spatial relationships and alleviate long-range dependency issues. The network was evaluated using the BraTS2021 dataset, with performance measured in terms of Dice coefficients and Hausdorff distances.

Results: DCEF-AVNet achieved Dice coefficients of 92.49%, 91.35%, and 91.96% for the whole tumor (WT), tumor core (TC), and enhancing tumor (ET) regions, respectively, significantly outperforming baseline methods. The model also demonstrated robust performance across multiple runs, with consistently low standard deviations in metrics.

Conclusions: The integration of dynamic convolution, enhanced feature fusion, and axial attention mechanisms enables DCEF-AVNet to deliver superior segmentation accuracy and robustness. These results underscore its potential for advancing automated brain tumor segmentation and improving clinical decision-making.

Purpose: Accurate depth estimation in surgical videos is a pivotal component of numerous image-guided surgery procedures. However, creating ground truth depth maps for surgical videos is often infeasible due to challenges such as inconsistent illumination and sensor noise. As a result, self-supervised depth and ego-motion estimation frameworks are gaining traction, eliminating the need for manually annotated depth maps. Despite the progress, current self-supervised methods still rely on known camera intrinsic parameters, which are frequently unavailable or unrecorded in surgical environments. We address this gap by introducing a self-supervised system capable of jointly predicting depth maps, camera poses, and intrinsic parameters, providing a comprehensive solution for depth estimation under such constraints.

Approach: We developed a self-supervised depth and ego-motion estimation framework, incorporating a cost volume-based auxiliary supervision module. This module provides additional supervision for predicting camera intrinsic parameters, allowing for robust estimation even without predefined intrinsics. The system was rigorously evaluated on a public dataset to assess its effectiveness in simultaneously predicting depth, camera pose, and intrinsic parameters.

Results: The experimental results demonstrated that the proposed method significantly improved the accuracy of ego-motion and depth prediction, even when compared with methods incorporating known camera intrinsics. In addition, by integrating our cost volume-based supervision, the accuracy of camera parameter estimation, including intrinsic parameters, was further enhanced.

Conclusions: We present a self-supervised system for depth, ego-motion, and intrinsic parameter estimation, effectively overcoming the limitations imposed by unknown or missing camera intrinsics. The experimental results confirm that the proposed method outperforms the baseline techniques, offering a robust solution for depth estimation in complex surgical video scenarios, with broader implications for improving image-guided surgery systems.

JMI Editor-in-Chief Bennett Landman discusses special issues and offers a few thoughts on the use of AI-assisted writing.

Purpose: Diverse population demographics can lead to substantial variation in the human anatomy. Therefore, standard anatomical atlases are needed for interpreting organ-specific analyses. Among abdominal organs, the pancreas exhibits notable variability in volumetric morphology, shape, and appearance, complicating the generalization of population-wide features. Understanding the common features of a healthy pancreas is crucial for identifying biomarkers and diagnosing pancreatic diseases.

Approach: We propose a high-resolution CT atlas framework optimized for the healthy pancreas. We introduce a deep-learning-based preprocessing technique to extract abdominal ROIs and leverage a hierarchical registration pipeline to align pancreatic anatomy across populations. Briefly, DEEDS affine and non-rigid registration techniques are employed to transfer patient abdominal volumes to a fixed high-resolution atlas template. To generate and evaluate the pancreas atlas, multi-phase contrast CT scans of 443 subjects (aged 15 to 50 years, with no reported history of pancreatic disease) were processed.

Results: The two-stage DEEDS affine and non-rigid registration outperforms other state-of-the-art tools, achieving the highest scores for pancreas label transfer across all phases (non-contrast: 0.497, arterial: 0.505, portal venous: 0.494, delayed: 0.497). External evaluation with 100 portal venous scans and 13 labeled abdominal organs shows a mean Dice score of 0.504. The low variance between the pancreases of registered subjects and the obtained pancreas atlas further illustrates the generalizability of the proposed method.

Conclusion: We introduce a high-resolution pancreas atlas framework to generalize healthy biomarkers across populations with multi-contrast abdominal CT. The atlases and the associated pancreas organ labels are publicly available through the Human Biomolecular Atlas Program (HuBMAP).