Journal of Medical Imaging最新文献

Purpose: The Response Evaluation Criteria in Solid Tumors (RECIST) relies solely on one-dimensional measurements to evaluate tumor response to treatments. However, thymic epithelial tumors (TETs), which frequently metastasize to the pleural cavity, exhibit a curvilinear morphology that complicates accurate measurement. To address this, we developed a physician-guided deep learning model and performed a retrospective study based on a patient cohort derived from clinical trials, aiming at efficient and reproducible volumetric assessments of TETs.

Approach: We used 231 computed tomography scans comprising 572 TETs from 81 patients. Tumors within the scans were identified and manually outlined to develop a ground truth that was used to measure model performance. TETs were characterized by their general location within the chest cavity: lung parenchyma, pleura, or mediastinum. Model performance was quantified on an unseen test set of 61 scans by mask Dice similarity coefficient (DSC), tumor DSC, absolute volume difference, and relative volume difference.

Results: We included 81 patients: 47 (58.0%) had thymic carcinoma; the remaining patients had thymoma B1, B2, B2/B3, or B3. The artificial intelligence (AI) model achieved an overall DSC of 0.77 per scan when provided with boxes surrounding the tumors as identified by physicians, corresponding to a mean absolute volume difference between the AI measurement and the ground truth of $16.1{\mathrm{cm}}^3$ and a mean relative volume difference of 22%.

Conclusion: We have successfully developed a robust annotation workflow and AI segmentation model for analyzing advanced TETs. The model has been integrated into the Picture Archiving and Communication System alongside RECIST measurements to enhance outcome assessments for patients with metastatic TETs.

Purpose: Medical image segmentation based on deep learning has played a crucial role in computer-aided medical diagnosis. However, they are still vulnerable to imperceptible adversarial attacks, which lead to potential misdiagnosis in clinical practice. Research on adversarial attack methods is beneficial for improving the robustness design of medical image segmentation models. Currently, there is a lack of research on adversarial attack methods toward deep learning-based medical image segmentation models. Existing attack methods often yield poor results in terms of both attack effects and image quality of adversarial examples and primarily focus on nontargeted attacks. To address these limitations and further investigate adversarial attacks on segmentation models, we propose an adversarial attack approach.

Approach: We propose an approach called momentum-driven adaptive feature-cosine-similarity with low-frequency constraint attack (MAFL-Attack). The proposed feature-cosine-similarity loss uses high-level abstract semantic information to interfere with the understanding of models about adversarial examples. The low-frequency component constraint ensures the imperceptibility of adversarial examples by constraining the low-frequency components. In addition, the momentum and dynamic step-size calculator are used to enhance the attack process.

Results: Experimental results demonstrate that MAFL-Attack generates adversarial examples with superior targeted attack effects compared with the existing Adaptive Segmentation Mask Attack method, in terms of the evaluation metrics of Intersection over Union, accuracy, $L_2$ , $L_{\infty }$ , Peak Signal to Noise Ratio, and Structure Similarity Index Measure.

Conclusions: The design idea of the MAFL-Attack inspires researchers to take corresponding defensive measures to strengthen the robustness of segmentation models.

Purpose: This study demonstrates the feasibility of using an LED array for hyperspectral imaging (HSI). The prototype validates the concept and provides insights into the design of future HSI applications. Our goal is to design, develop, and test a real-time, LED-based HSI prototype as a proof-of-principle device for in situ hyperspectral imaging using LEDs.

Approach: A prototype was designed based on a multiwavelength LED array and a monochrome camera and was tested to investigate the properties of the LED-based HSI. The LED array consisted of 18 LEDs in 18 different wavelengths from 405 nm to 910 nm. The performance of the imaging system was evaluated on different normal and cancerous ex vivo tissues. The impact of imaging conditions on the HSI quality was investigated. The LED-based HSI device was compared with a reference hyperspectral camera system.

Results: The hyperspectral signatures of different imaging targets were acquired using our prototype HSI device, which are comparable to the data obtained using the reference HSI system.

Conclusions: The feasibility of employing a spectral LED array as the illumination source for high-speed and high-quality HSI has been demonstrated. The use of LEDs for HSI can open the door to numerous applications in endoscopic, laparoscopic, and handheld HSI devices.

The editorial celebrates emerging breakthroughs and the foundational work that continues to shape the field.

Purpose: Mpox is a viral illness with symptoms similar to smallpox. A key clinical metric to monitor disease progression is the number of skin lesions. Manually counting mpox skin lesions is labor-intensive and susceptible to human error.

Approach: We previously developed an mpox lesion counting method based on the UNet segmentation model using 66 photographs from 18 patients. We have compared four additional methods: the instance segmentation methods Mask R-CNN, YOLOv8, and E2EC, in addition to a UNet++ model. We designed a patient-level leave-one-out experiment, assessing their performance using $F1$ score and lesion count metrics. Finally, we tested whether an ensemble of the networks outperformed any single model.

Results: Mask R-CNN model achieved an $F1$ score of 0.75, YOLOv8 a score of 0.75, E2EC a score of 0.70, UNet++ a score of 0.81, and baseline UNet a score of 0.79. Bland-Altman analysis of lesion count performance showed a limit of agreement (LoA) width of 62.2 for Mask R-CNN, 91.3 for YOLOv8, 94.2 for E2EC, and 62.1 for UNet++, with the baseline UNet model achieving 69.1. The ensemble showed an $F1$ score performance of 0.78 and LoA width of 67.4.

Conclusions: Instance segmentation methods and UNet-based semantic segmentation methods performed equally well in lesion counting. Furthermore, the ensemble of the trained models showed no performance increase over the best-performing model UNet, likely because errors are frequently shared across models. Performance is likely limited by the availability of high-quality photographs for this complex problem, rather than the methodologies used.

Purpose: Interventional MR imaging struggles with speed and efficiency. We aim to accelerate transrectal in-bore MR-guided biopsies for prostate cancer through undersampled image reconstruction and instrument localization by image segmentation.

Approach: In this single-center retrospective study, we used 8464 MR 2D multislice scans from 1289 patients undergoing a prostate biopsy to train and test a deep learning-based spatiotemporal MR image reconstruction model and a nnU-Net segmentation model. The dataset was synthetically undersampled using various undersampling rates ( $R=8$ , 16, 25, 32). An annotated, unseen subset of these data was used to compare our model with a nontemporal model and readers in a reader study involving seven radiologists from three centers based in the Netherlands. We assessed a maximum noninferior undersampling rate using instrument prediction success rate and instrument tip position (ITP) error.

Results: The maximum noninferior undersampling rate is 16-times for the temporal model (ITP error: 2.28 mm, 95% CI: 1.68 to 3.31, mean difference from reference standard: 0.63 mm, $P=.09$ ), whereas a nontemporal model could not produce noninferior image reconstructions comparable to our reference standard. Furthermore, the nontemporal model (ITP error: 6.27 mm, 95% CI: 3.90 to 9.07) and readers (ITP error: 6.87 mm, 95% CI: 6.38 to 7.40) had low instrument prediction success rates (46% and 60%, respectively) compared with the temporal model's 95%.

Conclusion: Deep learning-based spatiotemporal MR image reconstruction can improve time-critical intervention tasks such as instrument tracking. We found 16 times undersampling as the maximum noninferior acceleration where image quality is preserved, ITP error is minimized, and the instrument prediction success rate is maximized.

Purpose: Breast cancer remains a leading cause of death for women. Screening programs are deployed to detect cancer at early stages. One current barrier identified by breast imaging researchers is a shortage of labeled image datasets. Addressing this problem is crucial to improve early detection models. We present an active learning (AL) framework for segmenting breast masses from 2D digital mammography, and we publish labeled data. Our method aims to reduce the input needed from expert annotators to reach a fully labeled dataset.

Approach: We create a dataset of 1136 mammographic masses with pixel-wise binary segmentation labels, with the test subset labeled independently by two different teams. With this dataset, we simulate a human annotator within an AL framework to develop and compare AI-assisted labeling methods, using a discriminator model and a simulated oracle to collect acceptable segmentation labels. A UNet model is retrained on these labels, generating new segmentations. We evaluate various oracle heuristics using the percentage of segmentations that the oracle relabels and measure the quality of the proposed labels by evaluating the intersection over union over a validation dataset.

Results: Our method reduces expert annotator input by 44%. We present a dataset of 1136 binary segmentation labels approved by board-certified radiologists and make the 143-image validation set public for comparison with other researchers' methods.

Conclusions: We demonstrate that AL can significantly improve the efficiency and time-effectiveness of creating labeled mammogram datasets. Our framework facilitates the development of high-quality datasets while minimizing manual effort in the domain of digital mammography.

Purpose: Early-stage pneumonia is not easily detected, leading to many patients missing the optimal treatment window. This is because segmenting lesion areas from CT images presents several challenges, including low-intensity contrast between the lesion and normal areas, as well as variations in the shape and size of lesion areas. To overcome these challenges, we propose a segmentation network called DECE-Net to segment the pneumonia lesions from CT images automatically.

Approach: The DECE-Net adds an extra encoder path to the U-Net, where one encoder path extracts the features of the original CT image with the attention multi-scale feature fusion module, and the other encoder path extracts the contour features in the CT contour image with the contour feature extraction module to compensate and enhance the boundary information that is lost in the downsampling process. The network further fuses the low-level features from both encoder paths through the feature fusion attention connection module and connects them to the upsampled high-level features to replace the skip connections in the U-Net. Finally, multi-point deep supervision is applied to the segmentation results at each scale to improve segmentation accuracy.

Results: We evaluate the DECE-Net using four public COVID-19 segmentation datasets. The mIoU results for the four datasets are 80.76%, 84.59%, 84.41%, and 78.55%, respectively.

Conclusions: The experimental results indicate that the proposed DECE-Net achieves state-of-the-art performance, especially in the precise segmentation of small lesion areas.

Purpose: Alzheimer's disease (AD) is becoming more prevalent among the elderly, with projections indicating that it will affect a significantly large population in the future. Regardless of substantial research efforts and investments focused on exploring the underlying biological factors, a definitive cure has yet to be discovered. The currently available treatments are only effective in slowing disease progression if it is identified in the early stages of the disease. Therefore, early diagnosis has become critical in treating AD.

Approach: Recently, the use of deep learning techniques has demonstrated remarkable improvement in enhancing the precision and speed of automatic AD diagnosis through medical image analysis. We propose a hybrid model that integrates a convolutional variational auto-encoder (CVAE) with a vision transformer (ViT). During the encoding phase, the CVAE captures key features from the MRI scans, whereas the decoding phase reduces irrelevant details in MRIs. These refined inputs enhance the ViT's ability to analyze complex patterns through its multihead attention mechanism.

Results: The model was trained and evaluated using 14,000 structural MRI samples from the ADNI and SCAN databases. Compared with three benchmark methods and previous studies with Alzheimer's classification techniques, our approach achieved a significant improvement, with a test accuracy of 93.3%.

Conclusions: Through this research, we identified the potential of the CVAE-ViT hybrid approach in detecting minor structural abnormalities related to AD. Integrating unsupervised feature extraction via CVAE can significantly enhance transformer-based models in distinguishing between stages of cognitive impairment, thereby identifying early indicators of AD.

Purpose: Diabetic retinopathy (DR) is the leading cause of blindness among working-age adults globally. Although machine learning (ML) has shown promise for DR diagnosis, ensuring model generalizability requires training on data from diverse populations. Federated learning (FL) offers a potential solution by enabling model training on decentralized datasets. However, privacy concerns persist in FL due to potential privacy breaches, such as gradient inversion attacks, which can be used to reconstruct sensitive training data and may discourage participation from patients.

Approach: We developed and tested a computationally efficient FL framework that integrates homomorphic encryption (HE) to safeguard patient privacy using 6457 retinal fundus images from the APTOS-2019 and ODIR-5K datasets. First, features are extracted from distributed fundus images using RETFound, a large pretrained foundation model for retinal analysis. These encrypted features are then used to train a lightweight multiclass logistic regression head (MLRH) model for DR grade classification using FL.

Results: Experimental results show that the MLRH model trained using FL achieves similar performance compared with a fully fine-tuned RETFound model on centralized data, with the area under the receiver operating characteristic curve scores of $0.93\pm 0.01$ on APTOS-2019 and $0.78\pm 0.02$ on ODIR-5K. Efficiency improvements include a 95.9-fold reduction in computation time and a 63.0-fold reduction in data transfer needs compared with fine-tuning the full RETFound model with FL. In addition, results showed that integrating HE effectively protects patient data against gradient inversion attacks.

Conclusions: We advance privacy-preserving, ML-based DR screening technology, supporting the goal of equitable vision care worldwide.