Deep learning for cardiac CT segmentation for congenital heart disease: A systematic review Sinling Tiffany Yu, Zahra Amini, Shereen Fouad, Jan Novak, Antonio Fratini, Francesco Bardozzo Abstract PurposeDeep Learning (DL) has transformed cardiac image segmentation, yet its application to congenital heart disease (CHD) remains underexplored, with no prior systematic review in this emerging domain. This review evaluates current DL approaches for CHD CT segmentation, identifies best performing model architectures, assesses quality of reporting and highlights challenges for clinical implementation.Materials and methodsFollowing PRISMA guidelines, six databases: PubMed, IEEE Xplore, MDPI, Science Direct, Web of Science, and Scopus were searched (2004–2025). Extracted characteristics include source of dataset, ground truth labelling approach, DL architecture and evaluation methods. Reporting quality was evaluated using the Checklist for Artificial Intelligence in Medical imaging (CLAIM).Results14 studies met the inclusion criteria, reflecting the emergence of this research area. The UNet architecture and its variants dominated (8 studies). The highest reported Dice Similarity Coefficient (DSC) for cardiac structures segmentation was achieved utilising an advanced hybrid ResNet-based model (Aorta, DSC = 0.945). Well documented areas include the training approach (12 studies) and model description (all). Only two studies performed external validation, and two assessed inter- and intra-rater variability. No studies reported sample size determination and handing of missing data. Small vascular structures consistently underperformed compared to whole heart segmentation.ConclusionDeep learning models show strong potential for CHD CT segmentation but remain limited by small datasets, inconsistent reporting and lack of clinical evaluation. Advancing this field requires multidisciplinary collaboration, standardised reporting, and integration of clinical co- design. As the first comprehensive review in this area, we provide a strong baseline for evaluating future studies. 1. Introduction Congenital heart disease (CHD) is a chronic, life-long condition and the leading cause of mortality from birth defects worldwide []. It is defined as the structural abnormality of the heart or great vessels that occurs during foetal development. These morphological abnormalities lead to complications with cardiac function and circulation, often requiring early intervention and lifelong medical management []. Most CHD cases are diagnosed prenatally, through prenatal screening and foetal echocardiography; however, some children are born with undetected CHD and are only diagnosed postnatally. In critical cases where immediate intervention is necessary a prompt diagnosis is essential []. Surgical and interventional planning for CHD remains a challenge due to the small anatomy and unique nature of each malformation []. Hence, a personalised approach to diagnosis and surgical planning is essential for optimising patient outcomes. To support this personalised approach, detailed imaging of cardiac anatomy is essential. CT imaging is usually the modality of choice for evaluating CHD due to its rapid acquisition time and ability to produce high-resolution images. Image segmentation can be employed to accurately identify regions of interests within these images, and traditionally, is performed manually by a clinician or radiologist, but this is time consuming and subject to inter and intra-user variability []. Semi-automated segmentation tools which require minimal user input use methods such as thresholding and region-growing, combined with imaging properties and clinical knowledge to delineate structures and features in the image []. Growing research into segmentation and 3D reconstruction has demonstrated successful use of 3D models for surgical planning, teaching, and communication purposes [, ]. Recent studies highlight significant backlogs and pressure amongst radiologists due to increasing demand for imaging services []. To address these challenges, decision support systems have been proposed covering a broad range of applications such as clinical documentation, diagnostic support and risk prediction [, ]. Central to many of these decision support systems is Artificial Intelligence (AI). AI based models have become increasingly prominent in the field of biomedical research and clinical practice []. One significant approach is deep learning (DL), a subset of machine learning (ML) that involves the use of neural networks with multiple layers to automatically extract features from large datasets. This enables DL to effectively model complex patterns and relationships in data, making it suitable for a variety of tasks such as image segmentation, disease detection, and predictive analytics [, ]. This systematic review aims to provide an in-depth overview of current research on deep learning techniques for cardiac CT segmentation in congenital heart disease. Previous systematic reviews have investigated deep learning methods for cardiac image segmentation across various modalities, including studies focusing on classification of heart sounds [], echocardiograms [], and cardiac MRIs [, ]. Alnasser et al (2024) provided a comprehensive systematic review on deep learning methods for cardiac structure segmentation of CT images []. However, this primarily addressed cardiovascular diseases (CVD) in adults, which often cover acquired conditions rather than congenital heart disease. CHD presents unique challenges and differs significantly as it includes structural abnormalities present at birth, affecting cardiac function in ways that are distinct from adult-onset CVDs, and require specialised imaging and analysis techniques. Despite the importance of accurate cardiac CT segmentation for diagnosis and management of CHD, there is a lack of reviews focusing on this area. To the best of our knowledge, no review has specifically addressed deep learning techniques for cardiac CT segmentation in paediatric congenital heart disease. Therefore, we aim to address this gap through systematically reviewing current literature, summarising methodological trends, performance, and in addition provide a standardised reporting comparison of the selected studies using the most up-to-date version of the Checklist for Artificial Intelligence in Medical Imaging (CLAIM) []. Together, this provides a strong benchmark for current progress and guidance for further research and clinical translation. 2. Methods A systematic review was conducted in accordance with the Preferred Reporting Information for Systematic reviews and Meta-Analysis (PRISMA) guidelines. This review was not prospectively registered on PROSPERO. The full review protocol, including predefined search strategies and eligibility criteria was developed prior to conducting the review. Further details can be found in in supplementary materials. 2.1. Study selection A comprehensive search was conducted across six databases: PubMed, IEEE Xplore, MDPI, Science Direct, Web of Science, and Scopus including publications from 2004 to 2025. An example of the search string used for Web of Science is: ((ALL = Congenital heart disease) AND (ALL = CT) AND (ALL = Cardiac) AND (ALL = Segmentation) AND (ALL = Artificial Intelligence)) OR ((ALL = Congenital heart disease) AND (ALL = CT) AND (ALL = Cardiac) AND (ALL = Segmentation) AND (ALL = Machine learning)) OR ((ALL = Congenital heart disease) AND (ALL = CT) AND (ALL = Cardiac) AND (ALL = Segmentation) AND (ALL = Deep learning)). Full search strings and search strategy for all databases can be found in of supplementary materials. Endnote X9.3.3 was used to manage and screen references. Dual independent screening of the Title, Abstract and Full Text were performed by SY and AF, any disagreements were resolved through discussion and consensus. The last search was conducted 31st October 2025. This systematic review aims to answer the following questions: - -What type of DL models have been used for cardiac CT segmentation, and which models demonstrate better performance? - -How do current studies report study design and model performance? - -What are the current challenges and future research directions in integrating DL-based cardiac CT segmentation tools into clinical practice? 2.2. Eligibility criteria Both paediatric and paediatric-adult mixed studies were eligible, provided that congenital heart disease cases were included and segmentation of cardiac specific anatomy was described. Studies were not limited by specific CT acquisition protocols to enable assessment of model generalisability. Segmentation was defined as a required inclusion criterion with studies that combined segmentation with downstream tasks such as classification included as long as the segmentation performance was explicitly reported. No limits on minimum reporting metrics were applied to enable a comprehensive assessment of reporting quality. The full inclusion and exclusion criteria are outlined in . 2.3. Data extraction strategy A detailed data extraction was conducted dual independently by the main investigator (SY) and secondary investigator (AF), with any discrepancies resolved by discussion and consensus. The following extracted information were documented into a table: general paper descriptions (authors, title, year, country), dataset characteristics (source of dataset, imaging type, dataset size, CHD pathology type, dataset split) and research design (model architecture, segmentation software, structures segmented, ground truth labelling approach). Study authors were not contacted regarding missing or unclear information as we wanted to accurately document reporting practices. In cases where methodological details were not reported in the publication or supplementary materials, this was documented as “Not mentioned”. 2.4. Reporting and quality assessment Published literature often show variability and inconsistencies in documentation due to a lack of standardised reporting framework, making cross-study comparisons difficult. To address this, our systematic review uses the CLAIM 2024 checklist, a structured framework for documenting important aspects of AI studies, allowing for transparency, consistency, and reproducibility []. The final 14 studies were evaluated against 44 comprehensive criteria, covering aspects such as study design, dataset usage, data labelling, model training, evaluation methods and additional relevant factors. Five of the CLAIM criteria were excluded as they were deemed inapplicable or irrelevant for the purposes of comparison, resulting in 39 applicable items (Supplementary materials ). CLAIM scoring was conducted dual independently by two investigators (SY and ZA), any discrepancies were resolved by discussion and consensus. Adherence to CLAIM was scored in a binary manner (Yes/No) by referring to the definitions and guidance provided. Overall compliance was calculated as (number of items scored Yes/ 39 applicable items) x 100%. 2.5 Risk of bias Tools exist for assessing risk of bias in AI studies, but none are specifically designed for medical imaging segmentation tasks. QUADAS-2 developed in 2011 assesses diagnostic test accuracy whereas PROBAST+AI addresses prediction models [, ]. Deep learning segmentation models differ significantly as methodological bias is dependent on pixel and voxel level performance as well as annotation driven pipelines. Therefore, the PROBAST+AI framework was adapted to ensure the domain criteria’s address segmentation specific methodological considerations. Specifically, the ‘predictors’ domain which originally assessed clinical predictor variables, was reframed as ‘Input images and preprocessing’ to evaluate image acquisition and preprocessing pipelines, while the ‘Outcome’ domain was reframed as ‘Reference standard for segmentation’ to assess ground truth labelling, annotator qualifications, and inter-rater reliability. Signalling questions and performance metrics were also modified to reflect segmentation-specific considerations such as Dice Similarity Coefficient, Hausdorff distance and data leakage at the patient level. However, as this is a novel adaptation that has not been independently validated, the findings should be interpreted with this consideration in mind. The risk of bias assessment was performed dual independently by SY and ZA with disagreements resolved through discussion and consensus. The adapted template, including detailed instructions for implementation can be found in of supplementary materials. 3. Results From the initial search, 1388 articles were identified. Following removal of duplicates, title, abstract and full-text screening, 13 articles were selected for review. Additionally, 1 article was identified from secondary sources, totalling 14 articles for the final systematic review (). 3.1. Publication and Geographical trends Results show a diverse range of publication contributions from around the world with the highest number of publications originating from China (50%), followed by Korea and Japan (15%). The low number of publications per year highlights limited research in this area (). 3.2. Datasets 3.2.1. Datasets characteristics. The dataset size varied greatly between the 14 studies with the cohort size ranging from 20 to 3,750 participants. Ten studies used publicly available datasets [–] while the remaining conducted retrospective studies using data collected from their local hospitals [,–]. Pei et al (2023) incorporated datasets from two different hospital institutions allowing for cross-institutional validation. Nainamalai et al (2022) used a publicly available dataset for model training and validated on a local hospital dataset. This approach enhanced the model’s reliability by enabling evaluation on real clinical data (). None of the articles provided an explanation on how the sample size was determined and only four of the selected studies provided information on patient demographics [,,,]. Commonly reported demographics include the age, gender, and types of congenital heart defect present (). As with most clinical datasets, the ones used in the studies exhibited class imbalance, which can impact on model performance. For segmentation tasks, this imbalance can exist as disproportionate background regions to small foreground structures of interest. For classification tasks, imbalance often arises from unequal class distribution and representation of CHD types across the dataset ( in supplementary materials documents the class imbalance present in classification tasks). Despite the potential influence of class imbalance on model training and evaluation, only two studies addressed or tried to mitigate this. In details: Nainamalai et al (2022) recognised the heterogeneity of the public and local datasets. Difference in model performances were assessed using a permutation test score with significant variation found between the distributions of the public and local test set. Moreover, Pei et al (2023) addressed severe class imbalance by implementing over sampling, duplicating the under-represented class. Additionally, a novel classification loss (Recall-MCAM Loss) was used, which encourages the model to focus on correctly identifying the underrepresented class []. 3.2.2. Imaging methods. Within the selected studies, CT angiography was the most prevalent [–,,,]. Imaging parameters varied between hospital institutions, with no standard protocol used across all () – most studies using publicly available datasets did not mention key details on the imaging parameters. Xu et al (2023), investigated the use of multiple CT scanners to increase the generalisability of the results. Findings highlighted limited difference in performance loss (less than 2%) when the model was trained on data from one single machine and evaluated on others indicating good generalisation to new CT machines. Image preprocessing strategies were reported in 13 studies [–], common techniques include cropping, resampling, intensity normalisation as well as down sampling. Anish et al (2024) utilised the Scaled Functionality-Centric Contrast Limited Adaptive Histogram Equalisation (SF-CLAHE), that enhances image contrast by adjusting pixel values based on the histogram of pixel intensities []. Zhu et al (2025) incorporated a double-mix augmentation module which blends labelled and pseudo labelled CT volumes by category using weight distributions []. (Table documenting pre-processing strategies in ). 3.3. Segmentation and labelling 3.3.1. Investigated CHD types. CHD conditions covered across the studies varied. Xu et al (2023) included the highest number of CHD types, classifying 17 types of CHD. Anish et al (2024) included 16 different types of CHD in their study while both Yao et al (2023) and Barrowclough et al (2021) explored 14 types of CHD. 10 types of CHD were mentioned in the dataset Nainamalai et al (2022) used. The full list of CHD types mentioned across the papers is summarised in (). shows the dataset characteristics for the three open-source datasets mentioned across the 14 studies, with ImageCHD being most used. The ground truth labelling was reported in all studies. Source of ground truth labeller was described in 9 studies [,,–,–] as well as annotation tools used disclosed only across 5 studies, including: ITKSnap [], 3D Slicer [,], Materialise Mimics [], and Ziostation2 []. Two studies outlined strategies to minimise inter and intra variability during the ground truth labelling process [,]. The study by Xu et al (2023) implemented a multi-tiered approach; 17 resident physicians performed the initial segmentation, reviewed by 3 junior radiologists and any concerns further reviewed by a senior radiologist. In comparison, Nainamalai et al (2022) took on a 2-step approach, a cardiology fellow with 8 years of experience performed the segmentation which was then validated and approved by a paediatric cardiologist with 20 years of experience. Neither of the studies completed any quantitative assessments to evaluate variation between annotators. 3.4. Model characteristics, frameworks, and architectures Two types of model frameworks were present across the studies: single -task and multi-task. Single-task frameworks focus on solving one specific problem such as segmentation whereas multi-task framework models are developed for more than one task (i.e., segmentation and classification). The full table documenting types of model frameworks across the studies can be found in of the supplementary materials. Five main model architectures were identified across the 14 studies: UNet, CNN, RNN, GAN and Advanced hybrid-based models (Supplementary materials ). provides a summary of the deep learning methods across the studies. 3.4.1. UNet variants. The UNet is a popular model for biomedical image segmentation, it consists of a specialised encoder-decoder structure allowing the capture of high-resolution features as well as precise localisation []. Eight studies implemented UNet variants to enhance segmentation tasks [,,,–]. Yoshida et al (2021) implemented a two-stage approach, first was segmentation of the whole heart followed by a refined four-chamber segmentation. Barrowclough et al (2021) introduced a UNet implicit architecture using implicit spline functions to model the segmentation boundary, this significantly reduced the computational complexity and memory use with results outperforming traditional pixel-based prediction CNNs. Yao et al (2023) proposed a 2-stage framework, integrating clinical knowledge, findings documented improved segmentation accuracy. Xu et al (2023) further enhanced segmentation with a multi-task framework which mirrored clinical workflows, this achieved accuracy comparable to junior radiologists and reduced diagnosis time significantly. Liu et al (2024) used a UNet with gated self-attention mechanisms and a novel conditioned energy function module specifically designed for great vessel refinement, this saw a 1.05–1.42% improvement in DSC, IOU and HD95. 3.4.2. CNN variants. Convolutional Neural Networks (CNNs) are among the most popular architectures, known for their ability to capture hierarchical features. Park et al (2022) introduced a CNN-based contour and distance transform-guided network, the network enhances segmentation accuracy by focusing on boundary regions. Nainamalai et al (2022) used a DenseVNet to reduce manual segmentation time. Zhang et al (2025) used a semi-supervised double-mix framework with confidence-difficulty weights and cross-pseudo-supervision (CPS) co-training, this demonstrated that even with 5% labelled data, meaningful segmentation could still be achieved. 3.4.3. RNN variants. Recurrent Neural Networks (RNNs), designed for pattern recognition in sequential data was used by Anish et al (2024). A multi-task framework for segmentation and classification was proposed, this incorporated clustering and feature selection to optimise segmentation, followed by an inverse weight with logish normalised exponential function-based deep Jordan recurrent neural network (IW-LNEF-DJRNN) to classify 16 types of congenital heart disease. 3.4.4. GAN variants. Liao et al (2020) proposed a multi-modality transfer learning network (MMLNET) to improve MRI to CT segmentation. The advanced hybrid structure was composed of three main units: a GAN based generator to transform MRI images to CT like images, a CNN-based discriminator unit to differentiate between synthetic and real data and a UNet based segmentation unit. Findings demonstrated superior performance compared to traditional CNN and UNet-based models. 3.4.5. Advanced hybrid deep learning models. Pei et al (2023) adopted a multi-task learning framework, the hybrid model combined UNet-based segmentation with a CNN classification model to predict postoperative pulmonary venous obstruction (PPVO), improved outcomes were achieved over single-task approaches. Habijan et al (2021) developed a hybrid model which combined a ResNet-based encoder for deep feature extraction with a Variational Autoencoder (VAE) regularisation to capture latent representatives. The UNet inspired encoder-decoder structure with skip connections achieved DSC values comparable to state-of-the-art (SOTA) models. Zhu et al (2025) introduced a novel hybrid layer-based encoder-decoder framework based on Swin transformer elements. The approach integrates global volume mixing modules (Global contextual mixer) with local volume based multi-head attention mechanisms (Local contextual self-attention), findings showed increased DSC values compared to SOTA models. Li et al. (2025) introduced a two-stage hybrid segmentation framework. This combined a 3D UNet based architecture with transformer style multi-head attention and contour-aware feature guidance. The model achieved higher DSC and improved boundary accuracy compared to five other SOTA models. 3.5. Evaluation and performance metrics Segmentation performance was evaluated using pixel-based and overlap-based metrics (). Pixel-based metrics, which include Precision, Recall, Accuracy and Sensitivity assesses the performance at pixel level. Overlap-based metrics including the Dice Similarity Coefficient and Jaccard Index measures the agreement between the model prediction and the ground truth segmentation. Sensitivity was the most used pixel-based metric, and DSC was the most used overlap-based metric. Only a small proportion of studies tested their model on an external dataset to check its generalisability to unseen data [,]. Similarly, only 8 studies performed robustness and sensitivity analysis to assess performance on conditions such as the presence of outliers, noise, and artifacts in the dataset [,,–,]. 3.6. Segmentation performance The DSC across the 14 studies varied between the structures segmented (). Due to substantial heterogeneity across the studies including datasets, structures segmented and validation strategies, it is important to note that direct quantitative comparison of segmentation performance between models is not appropriate. The following reported performance metrics are summarised descriptively and should be interpreted within the context of each studies experimental setting. Two main approaches were taken, whole heart segmentation (which includes the entire blood pool of the four chambers and great vessels but without differentiations) and multi-structure segmentation (which distinguishes heart structures). Whole heart segmentation generally achieved higher reported DSC (value reaching 0.967 in one study), this is likely due to clearer anatomical boundaries. In contrast, smaller vessel structures such as the pulmonary veins (0.631), pulmonary arteries (0.671), SVC (0.521) and IVC (0.513) showcased a wider variability and range in reported DSC across studies, likely due to thinner walls and variable morphology making them harder to delineate, highlighting the need for improved training strategies. Habijan et al (2021) using a hybrid ResNet-based model, reported higher DSC for two structures (LA, Ao), however these results were obtained using a different dataset and validation strategies so should not be interpreted as directly outperforming other methods included in this review. To avoid misleading cross study comparisons, the complete table of quantitative results can be found in supplementary materials and , where performance is interpreted based on task description. 3.7. Risk of bias shows the risk of bias assessment for the 14 included studies based on the adapted PROBAST+AI framework. Model development and model evaluation were assessed separately using this framework. For model development, all studies were rated as unclear, mainly due to insufficient reporting of reference standard generation and analysis procedures. Domain specific assessment found ‘input images and preprocessing’ as the strongest domain with 12/14 studies (85.7%) reported as low risk [–,–,]. The ‘reference standard for segmentation’ and ‘analysis’ domains showed greatest methodological concern with 11/14 studies (78.6%) reported as unclear [–]. Due to most studies utilising open-source datasets, key concerns include the lack of transparent annotation protocols, demonstration of consistent reference standard labelling and evidence of sample size adequate for model complexity. Model evaluation demonstrated better methodological quality overall. Generally, low risk was reported across three domains (‘Participants and data sources’, ‘Input images and preprocessing’ and ‘reference standard for segmentation’). Three studies were reported low risk overall [,,], two were reported high risk [,] while nine were reported unclear [,–,]. 4. Discussion Research on AI and deep learning applications in cardiology has predominantly focused on adult populations, with paediatric CHD, especially segmentation methods using cardiac CT significantly underexplored []. This systematic review addresses this research gap by specifically evaluating deep learning approaches for CHD segmentation in paediatric cardiac CT, an emerging field evident by the limited number of studies meeting the inclusion criteria and all the publications being within the past five years. Across the 14 studies, several recurring strengths and gaps were identified that shape the current state of the field. Strengths include robust technical model performance, with UNet variants and advanced hybrid architectures achieving high segmentation performance for larger cardiac structures, alongside strong CLAIM compliance in domains such as model description, training approach, and performance metrics reporting. However, key gaps emerged in four areas: limited clinical evaluation and co-design with end users, inconsistent ground truth labelling practices, methodological and reporting limitations (particularly around sample size determination, handling of missing data, and terminology use), and underdeveloped explainability and interpretability strategies. These gaps are discussed in the following subsections. 4.1. Clinical evaluation and co-design Whilst technical model performance is essential, clinical utility of deep learning models requires both robust clinician evaluation and meaningful co-design with healthcare professionals. Only two studies included clinician evaluation [,], highlighting a persistent bias towards quantitative performance metrics over qualitative feedback from clinicians. This imbalance limits the understanding of DL model usability, integration into clinical workflows and interpretability of outputs for end users. Without clinician input, key factors such as trust, and decision-making support remain unexplored []. Although clinicians, radiologists, and medical professionals are often identified as the end users of these tools their involvement in many studies often stops following the ground truth labelling stage. Only two studies integrated clinical domain knowledge into their model design [,]. Future research requires adopting a balanced approach combining quantitative performance metrics with qualitative clinician feedback as well as strong clinical engagement through multidisciplinary collaboration ensuring that the outputs are meaningful and align with clinical needs, increasing the likelihood of clinical integration [,]. 4.2. Ground truth labelling The ground truth is considered the ‘Gold Standard’, and it is a significant factor influencing the model’s prediction []. Despite its importance, over half of the studies did not mention the annotation tool used [–] and five of them did not disclose their ground truth labelling source [,,–]. In addition, across AI and deep learning publications, there is limited research on validating ground truth, addressing labelling inter and intra variability and establishing standard protocols, such as using medically certified software. The term ‘expert’ is also subjective; with some studies using trained students followed by validation by a clinician due to the impracticality of having clinicians manually annotating large datasets. With varying levels of expertise, approaches such as consensus through majority voting might also include biases. 4.3. Methodology and reporting Our analyses revealed strong compliance to the CLAIM checklist (, ) across seven domains: use of AI, study aims, study goals, data sources, model description, performance metrics and internal evaluations. However, key areas require improvement. Notably, none of the reviewed studies mentioned how the intended sample size was determined (e.g., using power calculations) or how missing data was handled. Despite this being a commonly reported issue in DL studies, with the typical approach being that more data is better, this very much limits the rigour of the studies []. Model descriptions and training approaches were generally thorough providing clear documentation of software libraries, frameworks, training epochs and loss functions, the rationale for selecting hyperparameters was inadequately justified. With many reasons based on prior literature rather than experimental validation. Reporting of dataset demographics and clinical characteristics were also limited with the inclusion and exclusion criteria rarely provided. Although data splitting and assignment to partitions were well documented in eleven studies [,,,,,,], inconsistent use of terminology, such as the terms ‘validation’ and ‘testing’ used interchangeably compromises the reproducibility []. All 14 studies used patient-wise splitting, where all images from a given patient set are confined to a single partition (training, validation or testing). This approach is important to avoid information leakage where slices from one patient can appear in both training and testing sets, leading to biased performance estimates. The validation set is used during the model development process, to aid in fine-tuning of hyperparameters and prevent overfitting. Whereas the testing set is used after to evaluate the trained model’s performance on unseen images []. There should be no overlap between the training and testing datasets to prevent bias. Proper documentation and use of consistent terminology, especially when implementing niche or novel models, is essential in avoiding confusion and improving the reproducibility of studies. 4.4. Explainability and interpretability The use of quantitative performance metrics alone is insufficient, explainable approaches such as using masked class activation maps (MCAM), Gradient-weighted class activation maps (Grad-CAM) and attention maps were used in two studies [,]. While heatmap approaches enhance post-hoc explainability by indicating important areas for model predictions, interpretability is not necessarily improved for radiologists who are accustomed to viewing grayscale images in dark environments []. Quantification of explainability and interpretability remains a challenge due to their subjective nature and misalignments with clinical workflows may hinder understanding rather than aid in diagnosis. As interpretability is based on individual user needs and levels of expertise, a user-centric approach should be taken to ensure outputs are clinically relevant and interpretable for the end user. 5. Study limitations Our screening process did not incorporate a formal inter-reviewer agreement assessment such as Cohen’s Kappa, which may have introduced bias. This was because a consensus approach was employed where any disagreements between two reviewers were resolved through discussion until an agreement was reached. However, all screening decisions were reviewed independently by two investigators to minimise the risk of excluding relevant studies. This limitation should be taken into consideration when interpreting the findings. Additionally, this review was not prospectively registered on PROSPERO, however, the full predefined protocol is provided in supplementary materials . 6. Conclusion, future directions, and recommendations Our work has revealed promising advancements in the field, such as a growing use of deep learning approaches for CT segmentation for paediatric CHD, development of novel methods and advanced hybrid approaches that demonstrate superior segmentation performance as well as robust clinical evaluation methods that pave the way towards clinical adoption. However, significant areas for improvement were also identified. In more details, many studies critically under reported details on ground truth labelling, sample size determination, external dataset validation and integration of clinical expertise throughout the model development process. Furthermore, the limited availability of open-source datasets as well as significant dataset imbalance in which rarer complex CHD types are under-represented, collectively hinders advancement. In future, exploring generative AI techniques to augment training data with synthetic representations of rare CHD types may offer a promising strategy to mitigate class imbalance and improve model performance []. Addressing these gaps require adoption of standardised reporting protocols, rigorous experimental designs incorporating clinical domain knowledge and a multidisciplinary approach that actively involves clinicians not only in the ground truth labelling but validation and evaluation stages. Overall, through aligning deep learning model development with clinical needs and adopting a user-centric design, this can drive the development of robust, impactful, and clinically meaningful tools that enhance patient outcomes in congenital heart disease.