Journal of Medical Imaging最新文献

Purpose: Multiple instance learning (MIL) has emerged as the best solution for whole slide image (WSI) classification. It consists of dividing each slide into patches, which are treated as a bag of instances labeled with a global label. MIL includes two main approaches: instance-based and embedding-based. In the former, each patch is classified independently, and then, the patch scores are aggregated to predict the bag label. In the latter, bag classification is performed after aggregating patch embeddings. Even if instance-based methods are naturally more interpretable, embedding-based MILs have usually been preferred in the past due to their robustness to poor feature extractors. Recently, the quality of feature embeddings has drastically increased using self-supervised learning (SSL). Nevertheless, many authors continue to endorse the superiority of embedding-based MIL.

Approach: We conduct 710 experiments across 4 datasets, comparing 10 MIL strategies, 6 self-supervised methods with 4 backbones, 4 foundation models, and various pathology-adapted techniques. Furthermore, we introduce 4 instance-based MIL methods, never used before in the pathology domain.

Results: We show that with a good SSL feature extractor, simple instance-based MILs, with very few parameters, obtain similar or better performance than complex, state-of-the-art (SOTA) embedding-based MIL methods, setting new SOTA results on the BRACS and Camelyon16 datasets.

Conclusion: As simple instance-based MIL methods are naturally more interpretable and explainable to clinicians, our results suggest that more effort should be put into well-adapted SSL methods for WSI rather than into complex embedding-based MIL methods.

Purpose: We propose a deep learning framework, the cycle-guided denoising diffusion probability model (CG-DDPM), for cross-modality magnetic resonance imaging (MRI) synthesis. The CG-DDPM aims to generate high-quality MRIs of a target modality from an existing modality, addressing the challenge of missing MRI sequences in clinical practice.

Approach: The CG-DDPM employs two interconnected conditional diffusion probabilistic models, with a cycle-guided reverse latent noise regularization to enhance synthesis consistency and anatomical fidelity. The framework was evaluated using the BraTS2020 dataset, which includes three-dimensional brain MRIs with $T1$ -weighted, $T2$ -weighted, and FLAIR modalities. The synthetic images were quantitatively assessed using metrics such as multi-scale structural similarity measure (MSSIM), peak signal-to-noise ratio (PSNR), and mean absolute error (MAE). The CG-DDPM was benchmarked against state-of-the-art methods, including IDDPM, IDDIM, and MRI-cGAN.

Results: The CG-DDPM demonstrated superior performance across all cross-modality synthesis tasks (T1 → T2, T2 → T1, T1 → FLAIR, and FLAIR → T1). It consistently achieved the highest MSSIM values (ranging from 0.966 to 0.971), the lowest MAE (0.011 to 0.013), and competitive PSNR values (27.7 to 28.8 dB). Across all tasks, CG-DDPM outperformed IDDPM, IDDIM, and MRI-cGAN in most metrics and exhibited significantly lower uncertainty and inconsistency in MC-based sampling. Statistical analyses confirmed the robustness of CG-DDPM, with $p\text{-}\text{values}<0.05$ in key comparisons.

Conclusions: The proposed CG-DDPM provides a robust and efficient solution for cross-modality MRI synthesis, offering improved accuracy, stability, and clinical applicability compared with existing methods. This approach has the potential to streamline MRI-based workflows, enhance diagnostic imaging, and support precision treatment planning in medical physics and radiation oncology.

Purpose: Autism is one of the most common neurodevelopmental conditions, and it is characterized by restricted, repetitive behaviors and social difficulties that affect daily functioning. It is challenging to provide an early and accurate diagnosis due to the wide diversity of symptoms and the developmental changes that occur during childhood. We evaluate the feasibility of an explainable deep learning (DL) model using structural MRI (sMRI) to identify meaningful brain biomarkers relevant to autism in children and thus support its diagnosis.

Approach: A total of 452 $T1$ -weighted sMRI scans from children aged 9 to 11 years were obtained from the Autism Brain Imaging Data Exchange database. A DL model was trained to differentiate between autistic and typically developing children. Model explainability was assessed using saliency maps to identify key brain regions contributing to classification. Model performance was evaluated across 20 folds and compared with traditional machine learning models trained with regional volumetric features extracted from the sMRI scans.

Results: The model achieved a mean area under the receiver operating curve of 71.2%. The saliency maps highlighted brain regions that are known neuroanatomical and functional biomarkers associated with autism, such as the cuneus, pericalcarine, ventricles, lingual, vermal lobules, caudate, and thalamus.

Conclusions: We show the potential of interpretable DL models trained on sMRI data to aid in autism diagnosis within a narrowly defined pediatric age group. Our findings contribute to the field of explainable artificial intelligence methods in neurodevelopmental research and may help in clinical decision-making for autism and other neurodevelopmental conditions.

Purpose: The purpose of this study was to assess the variations in shear-wave speed (SWS) in individual thenar muscles under varied pinch forces in healthy adults. It was hypothesized that (1) SWS would vary among the individual thenar muscles, and (2) there would be an increase in SWS with increased pinch force.

Approach: Thirteen healthy participants' dominant hands were imaged using an ultrasound probe aligned longitudinally along the muscle fibers of the abductor pollicis brevis (APB), opponens pollicis (OPP), and flexor pollicis brevis (FPB). The SWS of each muscle was derived. Each participant completed trials consisting of randomly ordered pinch forces at 0, 10, and 20 N, 10% of maximum pinch force (MPF), and 20% MPF.

Results: The SWSs varied significantly among individual thenar muscles ( $p<0.01$ ) under absolute ( $p<0.01$ ) and relative forces ( $p<0.05$ ). There was a significant increase in SWS as the force increased from 0 to 20 N in the APB ( $p<0.001$ ) and OPP ( $p<0.001$ ), and not in the FPB ( $p=0.873$ ). There was a significant increase in SWS as the force increased from 0 to 20% MPF in the APB ( $p=0.005$ ), and not in the OPP ( $p=0.586$ ) or the FPB ( $p=0.984$ ).

Conclusions: The SWS of the APB and OPP increased as force increased and was different among the thenar muscles. This suggests SWS evaluations may be an appropriate method for evaluating muscles under tension, or different voluntary force conditions, specifically for the APB and OPP muscles.

Purpose: Weakly supervised learning (WSL) is widely used for histopathological image analysis by modeling images as sets of fixed-size patches and utilizing image-level diagnoses as weak labels. However, in multiclass classification scenarios, patches corresponding to a wide spectrum of diagnostic categories can co-exist in a single image, complicating the learning process. We aim to address label uncertainty in such multiclass settings.

Approach: We propose a two-branch architecture and a complementary training strategy to improve patch-based WSL. One branch estimates patch-level class likelihoods, whereas the other predicts per-class patch relevance weights. These outputs are combined into image-level class predictions via a relevance-weighted sum of per-patch class likelihoods. To further improve performance, we introduce a multilabel augmentation strategy that forms new training samples by combining patch sets and labels from pairs of images, resulting in multilabel samples that enrich the training set by increasing the chance of having more patches that are relevant to the augmented label sets.

Results: We evaluate our method on two challenging multiclass breast histopathology datasets for region of interest classification. The proposed architecture and training strategy outperform conventional weakly supervised methods, demonstrating improved classification accuracy and robustness, particularly in underrepresented classes.

Conclusions: The proposed architecture effectively models the complex relationship between image-level labels and patch-level content in multiclass histopathological image analysis. Combined with the image-level multilabel augmentation strategy, it improves learning under label uncertainty. These contributions hold potential for more accurate and scalable diagnostic support systems in digital pathology.

Purpose: We aim to investigate the performance of dedicated breast computed tomography (CT) for the detection of soft-tissue lesions and compare it to the detection of microcalcification clusters using cascaded systems analysis, with the intent of identifying which lesion type should be used for system optimization.

Approach: Signal and noise were propagated through the imaging chain using a cascaded systems model to obtain the modulation transfer function and noise power spectrum. Two imaging tasks were considered: a soft-tissue mass lesion modeled as a disk of 4 mm diameter and a cluster of microcalcifications modeled as calcium carbonate spheres of $220\mu m$ diameter. Detectability indices using three numerical observer models were obtained for various scintillator thicknesses and acquisition conditions at a fixed 4.5 mGy mean glandular dose.

Results: Detectability index trends are reversed between soft-tissue lesion and microcalcification cluster for the range of X-ray tube voltages and filtrations studied, indicating a potential need for compromise. However, for each of the 150 combinations studied (6 kV settings $\times 5$ Cu filter thicknesses $\times 5$ CsI:Tl scintillator thicknesses) and for each of the three numerical observer models, the detectability index for soft-tissue lesions always exceeded the microcalcification cluster.

Conclusion: When the lesion type is unknown, such as during breast cancer screening, it is more appropriate to optimize the system parameters for the task of detecting a microcalcification cluster, as the detectability index for the soft-tissue lesion exceeded that for the microcalcification cluster for all conditions investigated.

Purpose: X-ray detector noise decomposition and normalized noise power spectrum (NNPS) are two metrics proposed in the European Guidelines for the quality control (QC) of digital mammography (DM) systems. We aim to examine the reproducibility of these metrics in longitudinal testing and the relevance of the limiting values in the Guidelines and produce device-specific performance data.

Approach: Semiannual QC data for 55 DM systems (16 models, 6 manufacturers) were retrieved from our medical physics archive, giving a total of 455 QC tests covering a period of 5 years (10 tests). Average values of the detector response function fit coefficients, the electronic, quantum and structure noise coefficients, and the NNPS data were analyzed across DM models and longitudinally for a given device. The fraction of quantum noise at the clinical detector air kerma ( ${\mathrm{DAK}}_{50\mathrm{mm}}$ ) level was determined, along with the longitudinal change in NNPS.

Results: Coefficient of variation for the NNPS at $2.0{\mathrm{mm}}^{-1}$ was 0.04, averaged over all systems. Quantum noise evaluated at ${\mathrm{DAK}}_{50\mathrm{mm}}$ was the largest noise fraction for all devices studied, ranging from 61.2% to 98.2%. Electronic noise was generally lower for the latest X-ray detectors. Consequently, the fraction of quantum noise at ${\mathrm{DAK}}_{50\mathrm{mm}}$ has improved from 6.2% to 28.0% and corresponds to a broader quantum noise-limited range.

Conclusion: The evaluated noise metrics were reproducible, identified changes in X-ray detector performance, and have a useful role to play in QC testing. The average values have application as reference performance data.

Purpose: Whole slide image (WSI) classification relies on multiple instance learning (MIL) with spatial patch features, but current methods struggle to capture global dependencies due to the immense size of WSIs and the local nature of patch embeddings. This limitation hinders the modeling of coarse structures essential for robust diagnostic prediction.

Approach: We propose Fourier Transform Multiple Instance Learning (FFT-MIL), a framework that augments MIL with a frequency-domain branch to provide compact global context. Low-frequency crops are extracted from WSIs via the Fast Fourier Transform and processed through a modular FFT-Block composed of convolutional layers and Min-Max normalization to mitigate the high variance of frequency data. The learned global frequency feature is fused with spatial patch features through lightweight integration strategies, enabling compatibility with diverse MIL architectures.

Results: FFT-MIL was evaluated across six state-of-the-art MIL methods on three public datasets (BRACS, LUAD, and IMP). Integration of the FFT-Block improved macro $F1$ scores by an average of 3.51% and area under the curve by 1.51%, demonstrating consistent gains across architectures and datasets.

Conclusions: FFT-MIL establishes frequency-domain learning as an effective and efficient mechanism for capturing global dependencies in WSI classification, complementing spatial features and advancing the scalability and accuracy of MIL-based computational pathology. The source code is publicly available at https://github.com/irulenot/FFT-MIL.

Purpose: We aim to analyze higher-order textural components of digital breast tomosynthesis (DBT) images to quantify differences in the appearance of breast parenchyma produced by different vendors.

Approach: We included consecutive women who had normal screening DBT exams in January 2018 from a GE system and in adjacent years from Hologic systems. Laplacian fractional entropy (LFE), as a measure of non-Gaussian statistical properties of breast tissue texture, was calculated from for-presentation Craniocaudal (CC) view DBT slices and synthetic mammograms (SMs) through frequency-based filtering with Gabor filters, which were considered mathematical models for human visual response to image textures. The LFE values were compared within and across subjects and vendors along with secondary parameters (laterality, year-to-year, modality, and breast density) via two-way analysis of variance (ANOVA) tests using frequency as one of the two independent variables, and a $P$ -value $<0.05$ was considered statistically significant.

Results: A total of 8529 CC view DBT slices and SM images from 73 screening exams in 25 women were analyzed. Significant differences in LFE were observed for different frequencies ( $P<0.001$ ) and across vendors (GE versus Hologic DBT: $P<0.001$ , GE versus Hologic SM: $P<0.001$ ).

Conclusion: Significant differences in perception of breast parenchyma textures among two DBT vendors were demonstrated via higher-order non-Gaussian statistical properties. This finding extends previously observed differences in anatomical noise power spectra in DBT images and provides quantitative evidence to support caution in across-vendor comparative reading and will be beneficial to facilitate future development of vendor-neutral artificial intelligence algorithms for breast cancer screening.