All you need is data preparation: A systematic review of image harmonization techniques in Multi-center/device studies for medical support systems.

BACKGROUND AND OBJECTIVES: Artificial intelligence (AI) models trained on multi-centric and multi-device studies can provide more robust insights and research findings compared to single-center studies. However, variability in acquisition protocols and equipment can introduce inconsistencies that hamper the effective pooling of multi-source datasets. This systematic review evaluates strategies for image harmonization, which standardizes appearances to enable reliable AI analysis of multi-source medical imaging. METHODS: A literature search using PRISMA guidelines was conducted to identify relevant papers published between 2013 and 2023 analyzing multi-centric and multi-device medical imaging studies that utilized image harmonization approaches. RESULTS: Common image harmonization techniques included grayscale normalization (improving classification accuracy by up to 24.42 %), resampling (increasing the percentage of robust radiomics features from 59.5 % to 89.25 %), and color normalization (enhancing AUC by up to 0.25 in external test sets). Initially, mathematical and statistical methods dominated, but machine and deep learning adoption has risen recently. Color imaging modalities like digital pathology and dermatology have remained prominent application areas, though harmonization efforts have expanded to diverse fields including radiology, nuclear medicine, and ultrasound imaging. In all the modalities covered by this review, image harmonization improved AI performance, with increasing of up to 24.42 % in classification accuracy and 47 % in segmentation Dice scores. CONCLUSIONS: Continued progress in image harmonization represents a promising strategy for advancing healthcare by enabling large-scale, reliable analysis of integrated multi-source datasets using AI. Standardizing imaging data across clinical settings can help realize personalized, evidence-based care supported by data-driven technologies while mitigating biases associated with specific populations or acquisition protocols.

Development of a Deep Learning-Based System for Optic Nerve Characterization in Transorbital Ultrasound Images on a Multicenter Data Set.

OBJECTIVE: Characterization of the optic nerve through measurement of optic nerve diameter (OND) and optic nerve sheath diameter (ONSD) using transorbital sonography (TOS) has proven to be a useful tool for the evaluation of intracranial pressure (ICP) and multiple neurological conditions. We describe a deep learning-based system for automatic characterization of the optic nerve from B-mode TOS images by automatic measurement of the OND and ONSD. In addition, we determine how the signal-to-noise ratio in two different areas of the image influences system performance. METHODS: A UNet was trained as the segmentation model. The training was performed on a multidevice, multicenter data set of 464 TOS images from 110 subjects. Fivefold cross-validation was performed, and the training process was repeated eight times. The final prediction was made as an ensemble of the predictions of the eight single models. Automatic OND and ONSD measurements were compared with the manual measurements taken by an expert with a graphical user interface that mimics a clinical setting. RESULTS: A Dice score of 0.719 ± 0.139 was obtained on the whole data set merging the test folds. Pearson's correlation was 0.69 for both OND and ONSD parameters. The signal-to-noise ratio was found to influence segmentation performance, but no clear correlation with diameter measurement performance was determined. CONCLUSION: The developed system has a good correlation with manual measurements, proving that it is feasible to create a model capable of automatically analyzing TOS images from multiple devices. The promising results encourage further definition of a standard protocol for the automatization of the OND and ONSD measurement process using deep learning-based methods. The image data and the manual measurements used in this work will be available at 10.17632/kw8gvp8m8x.1.

Carotid Ultrasound Boundary Study (CUBS): Technical considerations on an open multi-center analysis of computerized measurement systems for intima-media thickness measurement on common carotid artery longitudinal B-mode ultrasound scans.

After publishing an in-depth study that analyzed the ability of computerized methods to assist or replace human experts in obtaining carotid intima-media thickness (CIMT) measurements leading to correct therapeutic decisions, here the same consortium joined to present technical outlooks on computerized CIMT measurement systems and provide considerations for the community regarding the development and comparison of these methods, including considerations to encourage the standardization of computerized CIMT measurements and results presentation. A multi-center database of 500 images was collected, upon which three manual segmentations and seven computerized methods were employed to measure the CIMT, including traditional methods based on dynamic programming, deformable models, the first order absolute moment, anisotropic Gaussian derivative filters and deep learning-based image processing approaches based on U-Net convolutional neural networks. An inter- and intra-analyst variability analysis was conducted and segmentation results were analyzed by dividing the database based on carotid morphology, image signal-to-noise ratio, and research center. The computerized methods obtained CIMT absolute bias results that were comparable with studies in literature and they generally were similar and often better than the observed inter- and intra-analyst variability. Several computerized methods showed promising segmentation results, including one deep learning method (CIMT absolute bias = 106 ± 89 µm vs. 160 ± 140 µm intra-analyst variability) and three other traditional image processing methods (CIMT absolute bias = 139 ± 119 µm, 143 ± 118 µm and 139 ± 136 µm). The entire database used has been made publicly available for the community to facilitate future studies and to encourage an open comparison and technical analysis (https://doi.org/10.17632/m7ndn58sv6.1).

Deep learning segmentation of transverse musculoskeletal ultrasound images for neuromuscular disease assessment.

Ultrasound imaging is a patient-friendly and robust technique for studying physiological and pathological muscles. An automatic deep learning (DL) system for the analysis of ultrasound images could be useful to support an expert operator, allowing the study of large datasets requiring less human interaction. The purpose of this study is to present a deep learning algorithm for the cross-sectional area (CSA) segmentation in transverse musculoskeletal ultrasound images, providing a quantitative grayscale analysis which is useful for studying muscles, and to validate the results in a large dataset. The dataset included 3917 images of biceps brachii, tibialis anterior and gastrocnemius medialis acquired on 1283 subjects (mean age 50 ± 21 years, 729 male). The algorithm was based on multiple deep-learning architectures, and its performance was compared to a manual expert segmentation. We compared the mean grayscale value inside the automatic and manual CSA using Bland-Altman plots and a correlation analysis. Classification in healthy and abnormal muscles between automatic and manual segmentation were compared using the grayscale value z-scores. In the test set, a Precision of 0.88 ± 0.12 and a Recall of 0.92 ± 0.09 was achieved. The network segmentation performance was slightly less in abnormal muscles, without a loss of discrimination between healthy and abnormal muscle images. Bland-Altman plots showed no clear trend in the error distribution and the two readings have a 0.99 Pearson's correlation coefficient (p < 0.001, test set). The ICC(A, 1) calculated between the z-score readings was 0.99. The algorithm achieves robust CSA segmentation performance and gives mean grayscale level information comparable to a manual operator. This could provide a helpful tool for clinicians in neuromuscular disease diagnosis and follow-up. The entire dataset and code are made available for the research community.

Carotid Ultrasound Boundary Study (CUBS): An Open Multicenter Analysis of Computerized Intima-Media Thickness Measurement Systems and Their Clinical Impact.

Common carotid intima-media thickness (CIMT) is a commonly used marker for atherosclerosis and is often computed in carotid ultrasound images. An analysis of different computerized techniques for CIMT measurement and their clinical impacts on the same patient data set is lacking. Here we compared and assessed five computerized CIMT algorithms against three expert analysts' manual measurements on a data set of 1088 patients from two centers. Inter- and intra-observer variability was assessed, and the computerized CIMT values were compared with those manually obtained. The CIMT measurements were used to assess the correlation with clinical parameters, cardiovascular event prediction through a generalized linear model and the Kaplan-Meier hazard ratio. CIMT measurements obtained with a skilled analyst's segmentation and the computerized segmentation were comparable in statistical analyses, suggesting they can be used interchangeably for CIMT quantification and clinical outcome investigation. To facilitate future studies, the entire data set used is made publicly available for the community at http://dx.doi.org/10.17632/fpv535fss7.1.

Automatic segmentation of ultrasound images of gastrocnemius medialis with different echogenicity levels using convolutional neural networks.

The purpose of this study was to develop an automatic method for the segmentation of muscle cross-sectional area on transverse B-mode ultrasound images of gastrocnemius medialis using a convolutional neural network(CNN). In the provided dataset images with both normal and increased echogenicity are present. The manually annotated dataset consisted of 591 images, from 200 subjects, 400 relative to subjects with normal echogenicity and 191 to subjects with augmented echogenicity. From the DICOM files, the image has been extracted and processed using the CNN, then the output has been post-processed to obtain a finer segmentation. Final results have been compared to the manual segmentations. Precision and Recall scores as mean ± standard deviation for training, validation, and test sets are 0.96 ± 0.05, 0.90 ± 0.18, 0.89 ± 0.15 and 0.97 ±0.03, 0.89± 0.17, 0.90 ± 0.14 respectively. The CNN approach has also been compared to another automatic algorithm, showing better performances. The proposed automatic method provides an accurate estimation of muscle cross-sectional area in muscles with different echogenicity levels.