Esmaeil Shahrak Nadimi

Background: AI has increasingly been used to generate synthetic images to overcome the scarcity of large, regulatory-compliant medical image datasets. However, their quality and actual utility for algorithm development across deep learning architectures, medical settings, and image types have been poorly assessed. Objective: This study aimed to review the use of synthetic images for algorithm development and to assess their quality and effects on downstream algorithms trained on these images, according to AI architecture, medical setting, and image type. Methods: The authors conducted a systematic review of studies published between January 2020 and September 2025. PubMed, Scopus, IEEE Xplore, and Embase were searched for peer-reviewed studies evaluating synthetic images generated from human medical imaging data. Data were extracted for imaging modalities, generative architectures, image-quality assessment methods, downstream model performance, explainability approaches, reproducibility practices, and risk of bias. Given substantial heterogeneity and incomplete statistical reporting, findings were synthesized descriptively. Results: A total of 160 studies were included. Generative adversarial network–based architectures dominated image generation (119 studies, 74.4%), although hybrid and diffusion-based approaches became increasingly represented in recent years. Image-quality evaluation was highly heterogeneous: 50 studies included human-reader evidence, and 101 reported an automated image-quality assessment. Synthetic-only training generally underperformed real-data training, including a median Dice similarity coefficient change of –10.2% for segmentation. Adding synthetic data to real training sets was associated with median improvements in classification precision (+7.8%), recall (+5.3%), F ₁ -score (+5.5%), accuracy (+4.3%), and area under the curve (+4.0%), while the median Dice coefficient improvement was +1.2%. Explainability evidence was reported in 68 (42.5%) studies and was rarely validated against expert annotations. Reproducibility reporting was inconsistent, and only 38% (60/160) of studies were judged to have low overall risk of bias. Conclusions: Unlike previous reviews restricted to particular diseases, imaging modalities, or generative architectures, this review provides a cross-domain quantitative synthesis of both synthetic-image quality assessment and downstream model performance. Synthetic images currently appear more valuable as a complement to real training data than as a replacement, with modest improvements in classification performance, smaller gains in segmentation, and generally poorer performance with synthetic-only training. For researchers, developers, and regulators, these findings highlight the need for standardized multimethod image-quality evaluation, complete reporting, and rigorous external validation before synthetic-image pipelines are translated into clinical AI applications.

Introduction: Increasing demand for colonoscopy continues to strain healthcare systems worldwide. Colon capsule endoscopy (CCE) offers a minimally invasive alternative, but its adoption is limited by high re-investigation rates. The aim of this study is to develop and evaluate clinical prediction models for selecting faecal immunochemical test (FIT) positive patients most suitable for CCE versus colonoscopy. Methods: We conducted a secondary analysis of data from CareForColon2015 randomized controlled trial (2020–2022), including individuals aged 50–74 years with a positive FIT. Logistic regression models were developed to predict CCE transit, bowel cleansing, completeness, and colonoscopy indication. Sixty candidate predictors were assessed, including demographics, lifestyle factors, FIT values, medications, perceived stress, and health literacy. Models were validated using repeated random subsampling and evaluated on a 10% hold-out set using the area under the receiver-operating-characteristic curve (AUC), Cohen’s K, and accuracy. Decision curve analysis (DCA) was performed to assess clinical utility. Results: CCE achieved complete transit in 92.1% and acceptable bowel cleansing in 71.3% of participants, with 69.6% of investigations deemed complete. Colonoscopy was indicated in 68.0% of cases, based on broad inclusion criteria, and 55.9%, based on more stringent criteria. Models predicting colonoscopy indication showed moderate performance (AUC 0.69–0.71; accuracy 65–67%; Cohen’s K 0.28–0.30). DCA indicated positive net benefit for both models within threshold probabilities of 0.5–0.75, supporting their potential to identify FIT-positive patients unlikely to benefit from immediate colonoscopy. Conclusions: Clinical prediction models may assist in post-FIT triage between CCE and colonoscopy. DCA suggests potential to reduce unnecessary colonoscopies by identifying low-risk patients suitable for initial CCE. External validation is needed before clinical implementation.

Objective: Colposcopy involves subjective visual assessment of cervical features that may indicate cervical dysplasia. Pattern recognition during colposcopy could be enhanced by artificial intelligence (AI). Using colposcopy images with precisely mapped multiple biopsy sites and corresponding histologic diagnoses, we developed an AI model, Cervix-AID-Net, to classify colposcopy images into low-grade disease [less than cervical intraepithelial neoplasia (CIN) grade 2] and high-grade disease (CIN grade 2 or above). The objective of this study was to compare the diagnostic performance of the Cervix-AID-Net model with the digital colposcope (DySIS) color map and colposcopists’ interpretations of the cervix in identifying low-grade and high-grade disease. Methods: The authors used 3,153 colposcopy images from 178 women, each with 4 biopsies, to train and validate the algorithm. Sensitivity, specificity, positive predictive value, negative predictive value, and accuracy were calculated with 95% CIs. Results: Cervix-AID-Net achieved a diagnostic accuracy of 99.8% (95% CI: 99.6–99.9) in classifying colposcopy images into low-grade and high-grade categories. This was significantly higher than the DySIS color map accuracy of 58.8% (95% CI: 51.1–66.1) and the accuracy of the colposcopist’s visual impression of the cervix (55.1%, 95% CI: 47.2%–62.5%). Conclusion: This first version of the Cervix-AID-Net demonstrated superior diagnostic accuracy compared with both the DySIS color map and colposcopists’ visual assessment. The results need confirmation in a prospective clinical trial.

BACKGROUND The use of artificial intelligence (AI) to generate synthetic images has been increasingly used to overcome the scarcity of large, regulatory-compliant medical images datasets. However, their quality and actual utility in algorithm development across deep learning architectures, medical settings, and image type has been poorly assessed. OBJECTIVE To revise the use of synthetic images for algorithm development and assess their quality and effect on downstream algorithms trained with these images, according to AI architecture, medical setting, and image type. METHODS We conducted a systematic review following PRISMA guidelines. PubMed was searched in January 2024 using terms related to synthetic image generation, diagnostic imaging, and model evaluation, restricted to January 2020-December 2023. Eligible studies evaluated diagnostic, or follow-up applications incorporating synthetic images, reported quantitative measures of image quality or model performance, used human imaging data, and were published as full-text English articles. Studies using synthetic images for non-diagnostic purposes, microscopic/histopathology imaging, and non-primary literature were excluded. Two independent reviewers conducted study selection with discrepancies resolved through group discussion. Data extraction focused on generative architectures, image quality metrics (fidelity, perceptual, Turing-style, distribution-alignment), and downstream model performance outcomes. Risk of bias was assessed using custom criteria based on reporting guidelines for observational studies. Due to insufficient variance reporting across studies, only descriptive analyses (and not pooled analyses) were conducted. RESULTS From 952 articles identified, 94 studies met inclusion criteria. MRI, X-ray, and CT were the predominant imaging modalities (representing the majority of studies), primarily applied in oncology, respiratory diseases, and neurology. GAN-based architectures dominated synthetic image generation (83% of studies). Among 49 distinct quality metrics identified, Turing-style realism metrics were most frequently used, followed by perceptual and fidelity metrics. Seventy-six studies evaluated downstream model performance, revealing that training exclusively on synthetic images often resulted in performance decrements, particularly for segmentation tasks (declining Dice scores). In contrast, augmenting original datasets with synthetic images showed more consistent improvements, especially for classification metrics across most imaging modalities. Risk of bias assessment revealed significant methodological concerns, with only 22% of studies rated as low risk; primary sources of bias included inadequate reporting of sampling strategies, population characteristics, and lack of measures of variance/confidence. CONCLUSIONS Synthetic images show promise for addressing dataset limitations in medical AI, but current evidence demonstrates substantial methodological heterogeneity and modest performance improvements that limit clinical translation. The field urgently requires standardized evaluation frameworks and consensus reporting guidelines. Future research should prioritize developing robust, clinically validated frameworks to guide responsible deployment of synthetic image-enhanced AI systems in healthcare. CLINICALTRIAL PROSPERO (CRD42023369125).

Machine unlearning has garnered significant attention due to its ability to selectively erase knowledge obtained from specific training data samples in an already trained machine learning model. This capability enables data holders to adhere strictly to data protection regulations. However, existing unlearning techniques face practical constraints, often causing performance degradation, demanding brief fine-tuning post unlearning, and requiring significant storage. In response, this paper introduces a novel class of machine unlearning algorithms. First method is partial amnesiac unlearning, integration of layer-wise pruning with amnesiac unlearning. In this method, updates made to the model during training are pruned and stored, subsequently used to forget specific data from trained model. The second method assimilates layer-wise partial-updates into label-flipping and optimization-based unlearning to mitigate the adverse effects of data deletion on model efficacy. Through a detailed experimental evaluation, we showcase the effectiveness of proposed unlearning methods. Experimental results highlight that the partial amnesiac unlearning not only preserves model efficacy but also eliminates the necessity for brief post fine-tuning, unlike conventional amnesiac unlearning. Moreover, employing layer-wise partial updates in label-flipping and optimization-based unlearning techniques demonstrates superiority in preserving model efficacy compared to their naive counterparts.

Semantic segmentation of biomedical images found its niche in screening and diagnostic applications. Recent methods based on deep learning convolutional neural networks have been very effective, since they are readily adaptive to biomedical applications and outperform other competitive segmentation methods. Inspired by the U-Net, we designed a deep learning network with an innovative architecture, hereafter referred to as AID-U-Net. Our network consists of direct contracting and expansive paths, as well as a distinguishing feature of containing sub-contracting and sub-expansive paths. The implementation results on seven totally different databases of medical images demonstrated that our proposed network outperforms the state-of-the-art solutions with no specific pre-trained backbones for both 2D and 3D biomedical image segmentation tasks. Furthermore, we showed that AID-U-Net dramatically reduces time inference and computational complexity in terms of the number of learnable parameters. The results further show that the proposed AID-U-Net can segment different medical objects, achieving an improved 2D F₁-score and 3D mean BF-score of 3.82% and 2.99%, respectively.

Gastrointestinal (GI) tract diseases are responsible for substantial morbidity and mortality worldwide, including colorectal cancer, which has shown a rising incidence among adults younger than 50. Although this could be alleviated by regular screening, only a small percentage of those at risk are screened comprehensively, due to shortcomings in accuracy and patient acceptance. To address these challenges, we designed an artificial intelligence (AI)-empowered wireless video endoscopic capsule that surpasses the performance of the existing solutions by featuring, among others: (1) real-time image processing using onboard deep neural networks (DNN), (2) enhanced visualization of the mucous layer by deploying both white-light and narrow-band imaging, (3) on-the-go task modification and DNN update using over-the-air-programming and (4) bi-directional communication with patient’s personal electronic devices to report important findings. We tested our solution in an in vivo setting, by administrating our endoscopic capsule to a pig under general anesthesia. All novel features, successfully implemented on a single platform, were validated. Our study lays the groundwork for clinically implementing a new generation of endoscopic capsules, which will significantly improve early diagnosis of upper and lower GI tract diseases.

For years, hepatologists have been seeking non-invasive methods able to detect significant liver fibrosis. However, no previous algorithm using routine blood markers has proven to be clinically appropriate in primary care. We present a novel approach based on artificial intelligence, able to predict significant liver fibrosis in low-prevalence populations using routinely available patient data. We built six ensemble learning models (LiverAID) with different complexities using a prospective screening cohort of 3352 asymptomatic subjects. 463 patients were at a significant risk that justified performing a liver biopsy. Using an unseen hold-out dataset, we conducted a head-to-head comparison with conventional methods: standard blood-based indices (FIB-4, Forns and APRI) and transient elastography (TE). LiverAID models appropriately identified patients with significant liver stiffness (> 8 kPa) (AUC of 0.86, 0.89, 0.91, 0.92, 0.92 and 0.94, and NPV ≥ 0.98), and had a significantly superior discriminative ability (p < 0.01) than conventional blood-based indices (AUC = 0.60–0.76). Compared to TE, LiverAID models showed a good ability to rule out significant biopsy-assessed fibrosis stages. Given the ready availability of the required data and the relatively high performance, our artificial intelligence-based models are valuable screening tools that could be used clinically for early identification of patients with asymptomatic chronic liver diseases in primary care.