Journal of Medical Imaging最新文献

Purpose: Accurate pixel-level segmentation is essential for medical image analysis, particularly in assisting diagnosis and treatment planning. However, fully supervised learning methods rely heavily on high-quality annotated data, which are often scarce due to the high cost of manual labeling, privacy concerns, and limited availability. We aim to reduce reliance on precise annotations and improve segmentation performance under weak supervision.

Approach: We propose scribble position and temporal contrast learning (SPTCL), an innovative segmentation method that combines contrastive learning with weak supervision. Our method leverages the spatial continuity in 3D medical image volumes and the anatomical similarities across different cardiac phases to construct a contrastive learning task for robust feature representation from unlabeled data. To enhance the feature extraction capabilities, we employ a pre-trained encoder, which is initially trained on the ACDC dataset using contrastive learning to capture robust feature representations. This pre-trained encoder is then transferred to a weakly supervised segmentation network with a dual-branch decoder for further fine-tuning on the task. The predictions from both branches are fused to generate refined pseudo-labels, which are iteratively used to guide network training with only coarse scribble annotations.

Results: Experiments on the ACDC dataset show that SPTCL outperforms existing models, achieving a Dice coefficient of 90.5%, with a 2.5% improvement over the baseline and a 1.7% improvement over the latest model. Furthermore, SPTCL reduces training time by $\sim 33\%$ .

Conclusions: SPTCL effectively addresses the challenges of limited annotation in medical image segmentation by uniting contrastive learning with weak supervision. It demonstrates strong potential for practical deployment in clinical settings where high-quality labels are difficult to obtain.

Purpose: Coronary artery disease is the leading global cause of mortality. Automated detection and scoring of calcified plaques can help cardiovascular risk assessment. We propose a deep learning method for automatic detection and scoring of coronary artery calcified plaques on noncontrast CT scans.

Approach: We utilized five datasets from one internal and four external tertiary care institutions, three of them with manually annotated plaques. A coronary artery calcified plaque detection model was developed using the state-of-the-art nnU-Net deep learning framework, incorporating simultaneous segmentation of the aorta, heart, and lungs to reduce false positives. The training data consisted of 641 noncontrast CT scans from three labeled datasets, representing diverse vascular disease etiologies. Agatston scores were automatically computed to quantify plaque burden. The model was tested on 160 labeled CT scans and compared with a previous detection method. In addition, Agatston scores were correlated with patient demographics and clinical outcomes using two unlabeled datasets.

Results: The predicted and reference Agatston scores demonstrated a strong correlation ( $r^2=0.973$ ), with a precision of 89.3%, recall of 89.1%, and an average Dice score of $75.0\pm 16.0\%$ on the labeled testing datasets. The stratified four Agatston groups achieved 92.0% accuracy and a Cohen's Kappa of 0.913. In the unlabeled datasets, Agatston groups showed significant correlations with the Framingham risk score, cardiovascular disease, heart failure, cancer status, fragility fracture risk, smoking, and age, whereas remaining consistent across race and scanner types.

Conclusions: Coronary artery plaques were accurately detected and segmented using the proposed nnU-Net-based method on noncontrast CT scans. The Agatston-score-based plaque burden assessment facilitates cardiovascular risk stratification, enabling opportunistic screening and population-based studies.

Purpose: To avoid over-treatment of prostate cancer patients following screening for elevated prostate-specific antigen levels, keeping patients on active surveillance has been suggested as an alternative to radical treatment. This means recurring visits for patients with low-grade cancer to monitor progression. Our aim was to develop an artificial intelligence-based model that can identify high-risk patients in a cohort of prostate cancer patients on active surveillance.

Approach: We have developed a multiple instance learning-based framework for predicting the longitudinal outcomes for prostate cancer patients on active surveillance. Our models were trained only on whole-slide images with patient-level labels without using explicit Gleason grades. We employed the UNI-2 foundation model and the well-established attention-based multiple instance learning approach. We further evaluated our models by fitting Cox proportional hazards models and testing them on an external dataset.

Results: With this approach, we achieved an average area under the receiver operator characteristic curve of 0.958 (95% CI, 0.957 to 0.959). Fitting Cox models to the predicted probabilities achieved a $C$ -index of 0.824 and a hazard ratio of 2.32. However, all models showed a large drop in performance when evaluated on an external dataset.

Conclusion: We show that avoiding Gleason grades is beneficial for longitudinal outcome prediction of prostate cancer. Our results suggest that benign prostate tissue contains prognostic information. However, before our models could be used clinically, much more work remains to improve the generalization.

Purpose: The scarcity of high-quality curated labeled medical training data remains one of the major limitations in applying artificial intelligence systems to breast cancer diagnosis. Deep models for mammogram analysis and mass (or micro-calcification) detection require training with a large volume of labeled images, which are often expensive and time-consuming to collect. To reduce this challenge, we proposed a method that leverages self-supervised learning (SSL) and a deep hybrid model, named HybMNet, which combines local self-attention and fine-grained feature extraction to enhance breast cancer detection on screening mammograms.

Approach: Our method employs a two-stage learning process: (1) SSL pretraining: We utilize Efficient Self-Supervised Vision Transformers, an SSL technique, to pretrain a Swin Transformer (Swin-T) using a limited set of mammograms. The pretrained Swin-T then serves as the backbone for the downstream task. (2) Downstream training: The proposed HybMNet combines the Swin-T backbone with a convolutional neural network (CNN)-based network and a fusion strategy. The Swin-T employs local self-attention to identify informative patch regions from the high-resolution mammogram, whereas the CNN-based network extracts fine-grained local features from the selected patches. A fusion module then integrates global and local information from both networks to generate robust predictions. The HybMNet is trained end-to-end, with the loss function combining the outputs of the Swin-T and CNN modules to optimize feature extraction and classification performance.

Results: The proposed method was evaluated for its ability to detect breast cancer by distinguishing between benign (normal) and malignant mammograms. Leveraging SSL pretraining and the HybMNet model, it achieved an area under the ROC curve of 0.864 (95% CI: 0.852, 0.875) on the Chinese Mammogram Database (CMMD) dataset and 0.889 (95% CI: 0.875, 0.903) on the INbreast dataset, highlighting its effectiveness.

Conclusions: The quantitative results highlight the effectiveness of our proposed HybMNet and the SSL pretraining approach. In addition, visualizations of the selected region of interest patches show the model's potential for weakly supervised detection of microcalcifications, despite being trained using only image-level labels.

Purpose: Deep generative models and synthetic data generation have become essential for advancing computer-assisted diagnosis and treatment. We explore one such emerging and particularly promising application of deep generative models, namely, the generation of virtual contrast enhancement. This allows to predict and simulate contrast enhancement in breast magnetic resonance imaging (MRI) without physical contrast agent injection, thereby unlocking lesion localization and categorization even in patient populations where the lengthy, costly, and invasive process of physical contrast agent injection is contraindicated.

Approach: We define a framework for desirable properties of synthetic data, which leads us to propose the scaled aggregate measure (SAMe) consisting of a balanced set of scaled complementary metrics for generative model training and convergence evaluation. We further adopt a conditional generative adversarial network to translate from non-contrast-enhanced $T1$ -weighted fat-saturated breast MRI slices to their dynamic contrast-enhanced (DCE) counterparts, thus learning to detect, localize, and adequately highlight breast cancer lesions. Next, we extend our model approach to jointly generate multiple DCE-MRI time points, enabling the simulation of contrast enhancement across temporal DCE-MRI acquisitions. In addition, three-dimensional U-Net tumor segmentation models are implemented and trained on combinations of synthetic and real DCE-MRI data to investigate the effect of data augmentation with synthetic DCE-MRI volumes.

Results: Conducting four main sets of experiments, (i) the variation across single metrics demonstrated the value of SAMe, and (ii) the quality and potential of virtual contrast injection for tumor detection and localization were shown. Segmentation models (iii) augmented with synthetic DCE-MRI data were more robust in the presence of domain shifts between pre-contrast and DCE-MRI domains. The joint synthesis approach of multi-sequence DCE-MRI (iv) resulted in temporally coherent synthetic DCE-MRI sequences and indicated the generative model's capability of learning complex contrast enhancement patterns.

Conclusions: Virtual contrast injection can result in accurate synthetic DCE-MRI images, potentially enhancing breast cancer diagnosis and treatment protocols. We demonstrate that detecting, localizing, and segmenting tumors using synthetic DCE-MRI is feasible and promising, particularly considering patients where contrast agent injection is risky or contraindicated. Jointly generating multiple subsequent DCE-MRI sequences can increase image quality and unlock clinical applications assessing tumor characteristics related to its response to contrast media injection as a pillar for personalized treatment planning.

The editorial introduces the JMI Special Issue on Advances in Breast Imaging, reflecting on the current forefront of breast imaging research.

Purpose: Biliary tract cancer, also known as intrahepatic cholangiocarcinoma (IHCC), is a rare disease that shows no clear symptoms during its early stage, but its prognosis depends highly on the cancer subtype. Hence, an accurate cancer subtype classification model is necessary to provide better treatment plans to patients and to reduce mortality. However, annotating histopathology images at the pixel or patch level is time-consuming and labor-intensive for giga-pixel whole slide images. To address this problem, we propose a weakly supervised method for classifying IHCC subtypes using only image-level labels.

Approach: The core idea of the proposed method is to detect regions (i.e., subimages or patches) commonly included in all subtypes, which we name the "hidden class," and to remove them via iterative application of contrastive loss and label smoothing. Doing so will enable us to obtain only patches that faithfully represent each subtype, which are then used to train the image-level classification model by multiple instance learning (MIL).

Results: Our method outperforms the state-of-the-art weakly supervised learning methods ABMIL, TransMIL, and DTFD-MIL by $\sim 17\%$ , 18%, and 8%, respectively, and achieves performance comparable to that of supervised methods.

Conclusions: The introduction of a hidden class to represent patches commonly found across all subtypes enhances the accuracy of IHCC classification and addresses the weak labeling problem in histopathology images.

Purpose: Various deep learning (DL) approaches have been developed for estimating scatter radiation in digital breast tomosynthesis (DBT). Existing DL methods generally employ an end-to-end training approach, overlooking the underlying physics of scatter formation. We propose a deep learning approach inspired by asymmetric scatter kernel superposition to estimate scatter in DBT.

Approach: We use the network to generate the scatter amplitude distribution as well as the scatter kernel width and asymmetric factor map. To account for variations in local breast thickness and shape in DBT projection data, we integrated the Euclidean distance map and projection angle information into the network design for estimating the asymmetric factor.

Results: Systematic experiments on numerical phantom data and physical experimental data demonstrated the outperformance of the proposed approach to UNet-based end-to-end scatter estimation and symmetric kernel-based approaches in terms of signal-to-noise ratio and structure similarity index measure of the resulting scatter corrected images.

Conclusions: The proposed method is believed to have achieved significant advancement in scatter estimation of DBT projections, allowing a robust and reliable physics-informed scatter correction.

Purpose: We examine the transformative potential of artificial intelligence (AI) in medical imaging diagnosis, focusing on improving diagnostic accuracy and efficiency through advanced algorithms. It addresses the significant challenges preventing immediate clinical adoption of AI, specifically from technical, ethical, and legal perspectives. The aim is to highlight the current state of AI in medical imaging and outline the necessary steps to ensure safe, effective, and ethically sound clinical implementation.

Approach: We conduct a comprehensive discussion, with special emphasis on the technical requirements for robust AI models, the ethical frameworks needed for responsible deployment, and the legal implications, including data privacy and regulatory compliance. Explainable artificial intelligence (XAI) is examined as a means to increase transparency and build trust among healthcare professionals and patients.

Results: The analysis reveals key challenges to AI integration in clinical settings, including the need for extensive high-quality datasets, model reliability, advanced infrastructure, and compliance with regulatory standards. The lack of explainability in AI outputs remains a barrier, with XAI identified as crucial for meeting transparency standards and enhancing trust among end users.

Conclusions: Overcoming these barriers requires a collaborative, multidisciplinary approach to integrate AI into clinical practice responsibly. Addressing technical, ethical, and legal issues will support a softer transition, fostering a more accurate, efficient, and patient-centered healthcare system where AI augments traditional medical practices.