The rapid proliferation of artificial intelligence in medical imaging is currently hindered by a significant disconnect between high-performing research models and the rigorous demands of clinical environments. Key challenges include data interoperability issues between research formats and clinical standards, hardware dependencies that limit portability, and the opaque “black-box” nature of deep learning models which erodes clinician trust. In this work, we propose a comprehensive intelligent information system designed to bridge this gap by unifying standards-compliant data ingestion, accelerated inference, and knowledge-infused reasoning into a single auditable workflow. Our approach integrates a robust DICOM and NIfTI ingestion pipeline with built-in anonymization, a hardware-agnostic ONNX inference engine, and a novel graph-based classification module that explicitly models anatomical relationships. Evaluated on the public ACDC benchmark, the proposed system demonstrates superior performance, with the segmentation module achieving a mean Dice Similarity Coefficient of 0.939 and the knowledge-integrated classifier attaining a diagnostic accuracy of 94.0%. The significant conclusion of this study is that by systematically integrating privacy controls, hardware portability, and graph-based knowledge representation, it is possible to create a deployment-ready AI blueprint that is both scientifically reproducible and clinically trustworthy.
The field of artificial intelligence (AI) for medical imaging has witnessed exponential growth in
recent years, driven by the advent of deep learning architectures that often surpass human-level
performance in specific diagnostic tasks. However, a substantial chasm remains between the
experimental success of these models in controlled research environments and their practical
utility in real-world clinical settings [
Furthermore, the deployment landscape is complicated by hardware heterogeneity. Research
models are typically trained on high-end NVIDIA GPUs using frameworks like PyTorch, but
clinical workstations vary widely in their computational capabilities, ranging from standard CPUs
to GPUs from different vendors. This necessitates an inference strategy that is both portable and
performant. The Open Neural Network Exchange (ONNX) format and its associated Runtime
engine offer a solution by providing an intermediate representation that can be executed across
diverse hardware backends [
Perhaps the most critical barrier to adoption is the “black-box” nature of modern deep neural
networks. In high-stakes medical decision-making, accuracy alone is insufficient; clinicians require
transparency and justification for algorithmic predictions. While purely data-driven models like the
U-Net [
The problem under consideration is the absence of a unified, end-to-end framework that
systematically addresses these disparate requirements, i.e., standards compliance, hardware
portability, and knowledge-infused reasoning [
In this work, we present a novel scientific contribution by architecting an intelligent information system that integrates these components from the ground up. By combining a standards-compliant ingestion module, a portable ONNX-based segmentation engine, and a graph convolutional network (GCN) for structured reasoning, we provide a holistic solution to the deployment gap.
The goal of this study is to improve knowledge integration fidelity and downstream reasoning accuracy by unifying standards-compliant ingestion, portable ONNX inference, and graphstructured classification with calibration-aware evaluation. To achieve this goal, we present three major contributions: 1.
A complete system architecture that spans DICOM/NIfTI ingestion with built-in anonymization, accelerated ONNX inference, volumetric segmentation, and manifest-driven data export for full reproducibility.
A novel graph-based classification module, KI-GCN, derived from GCNs, which aggregates structured features from segmentation masks and patient metadata to enhance diagnostic reasoning. We also specify an optional multi-teacher knowledge distillation objective for deploying compressed, efficient models.
A deployment-oriented evaluation protocol that includes standard segmentation metrics, probabilistic classification metrics, and critical calibration diagnostics like reliability diagrams to ensure model trustworthiness.
The remainder of this paper is organized as follows. Section 2 reviews the state of the art in medical imaging interoperability, segmentation architectures, and knowledge integration methods. Section 3 details the proposed system architecture, including the formalization of the ingestion, segmentation, and graph-based classification modules. Section 4 presents the experimental results on the ACDC and M&Ms-2 datasets, providing a comparative analysis against state-of-the-art methods. Section 5 analyzes the implications of these findings, system throughput, and limitations. Finally, Section 6 summarizes the contributions and outlines future directions.
Our research is situated at the intersection of interoperability standards, hardware acceleration, advanced deep learning architectures for segmentation, and methods for knowledge integration. This section reviews the state of the art in these domains to contextualize the proposed intelligent information system.
The foundation of any clinical AI system is its ability to handle standardized data formats. The
DICOM standard serves as the global lingua franca for medical imaging, with its Part 1 (PS3.1)
defining the overarching structure and semantic interoperability requirements for clinical data
exchange [
Regarding model deployment and hardware acceleration, the ONNX Runtime engine has
emerged as a critical technology for ensuring portability. It abstracts the execution of model graphs
through a system of pluggable Execution Providers (EPs), allowing the same model file to run
efficiently on CPUs, NVIDIA GPUs via CUDA [13], and Windows-based GPUs via DirectML [14].
This flexibility is essential for clinical environments where hardware specifications cannot be
guaranteed. Recent comparative analyses have highlighted the necessity of such acceleration
frameworks to reduce inference latency and computational overhead in production settings [
In the domain of medical image segmentation, the U-Net architecture remains the cornerstone,
featuring a symmetric encoder-decoder structure with skip connections that preserve spatial
information [
To move beyond the “black-box” paradigm, integrating explicit knowledge is crucial. Neural networks, particularly GCNs, provide a mathematical framework for modeling anatomical structures as interconnected nodes, allowing for reasoning based on spatial and functional relationships rather than just pixel intensities [16]. Advanced variants like graph attention networks have further refined this approach by learning to weigh the importance of different anatomical connections [17]. Additionally, knowledge distillation offers a pathway to compress these complex reasoning capabilities into lightweight models suitable for deployment, transferring insights from large “teacher” ensembles to efficient “student” models [18, 19]. Recent work in our group has extended these concepts to adaptive multi-teacher distillation strategies, enhancing robustness against domain shifts [20, 21].
Finally, trust in AI systems is predicated not just on accuracy, but on calibration, i.e., the alignment between predicted confidence and actual correctness. Methods such as reliability diagrams and temperature scaling are essential for diagnosing and correcting miscalibration [22]. Emerging techniques like proximity-informed calibration continue to push the boundaries of model reliability [23].
The primary objective of this study is to synthesize these diverse technological threads into a single, cohesive system. The main tasks to fulfill this objective are: (i) to design and implement a modular software architecture for the end-to-end medical imaging workflow, (ii) to develop and integrate a graph-based reasoning module that leverages segmentation outputs for improved classification, and (iii) to validate the entire system’s performance and reproducibility on public benchmark datasets.
We formalize the proposed intelligent information system as a sequence of interconnected processing modules that execute a single, manifest-driven workflow. The system is designed to transform raw medical imaging data into actionable, explainable diagnostic insights. Detailed implementation specifics and user manuals are provided in the accompanying technical report [24]. In this section, we define the mathematical formulations and algorithmic logic underpinning the core components: ingestion, segmentation, and graph-based knowledge integration.
Let a dataset be denoted by ={(V i , M i , Di)}iN=1, where for each of N patients, V i represents the input medical image volume (e.g., a cardiac MRI series), M i represents the ground-truth anatomical segmentation mask, and Di represents the associated clinical diagnosis or classification label. The system architecture, illustrated in Figure 1, processes these inputs through four distinct stages: ingestion, segmentation, knowledge graph construction, and classification.
The system processes each patient’s data i through a sequential pipeline formalized below.
The ingestion module is responsible for the secure and accurate loading of medical data. It utilizes
the FO-DICOM library to parse DICOM series, ensuring that all slices are ordered correctly based
on the ‘ImagePositionPatient’ (0020,0032) tag. To adhere to privacy regulations (e.g., GDPR,
HIPAA), the module implements a configurable anonymization engine compliant with the DICOM
PS3.15 Basic Profile. Identifiers such as ‘PatientName’ and ‘PatientID’ are hashed or removed [
For research data in NIfTI format, the system parses the affine header to normalize voxel
spacing and reorient the volume to the canonical RAS coordinate system [
Synergistic Knowledge-Integrated Framework for Segmentation (SKIF-Seg) is the system’s segmentation engine, designed for hardware portability via ONNX Runtime. The module accepts the preprocessed volume V ′i and predicts a dense probability map Pi. The inference process is abstracted to support multiple backends:
CPU Execution Provider: Uses MKLDNN/OpenBLAS for optimized execution on standard processors.
CUDA Execution Provider: Leverages NVIDIA’s cuDNN and TensorRT libraries for highthroughput GPU inference [13].
DirectML Execution Provider: Provides vendor-agnostic GPU acceleration on Windows, supporting AMD, Intel, and NVIDIA hardware [14].
The output Pi∈[
To incorporate anatomical reasoning, we introduce the Knowledge Integration Graph Convolutional Network (KI-GCN). We define a graph G=(V , E ) where nodes V correspond to segmented structures (e.g., Left Ventricle, Myocardium, Right Ventricle) and edges E encode spatial adjacency and functional connectivity.
For each node v ∈V , we compute a feature vector v derived from the segmentation mask M^ i, including volume, surface area, sphericity, and centroid displacement. The graph is processed using spectral graph convolution layers defined by the propagation rule as follows
H (l+1)=σ ( D~−12 ~A D~−12 H (l) W (l)), (1) ~ ~ where H (ℓ) is the feature matrix at layer ℓ, A is the adjacency matrix with self-loops, D is the degree matrix, and W (ℓ) is the learnable weight matrix [16].
This process allows the model to learn features that depend on the structural configuration of the heart, rather than treating geometry as a flat vector. The final node embeddings are pooled to form a global graph representation G, which is classified into diagnostic categories.
To enable efficient deployment on edge devices, we employ a multi-teacher knowledge distillation strategy. The training objective combines the standard cross-entropy loss with a distillation term that aligns the student’s logits z(s) with the soft targets from an ensemble of teacher models z¯(t) as presented below
L=α LCE( y , softmax ( z(s)))+(1−α ) τ 2 KL(softmax( z¯(t) τ )|| softmax( z(s) τ (2) where τ is the temperature parameter controlling the softness of the probability distributions, and α balances the two loss components [18, 19].
Our evaluation protocol is designed to be comprehensive, deployment-oriented, and fully reproducible.
For segmentation performance, let X and Y be the predicted and ground-truth masks, respectively.
We quantify overlap using the Dice Similarity Coefficient (DSC) [25], as follows Additionally, we calculate the Jaccard Index (IoU) [26], defined in Equation 4:
For boundary accuracy, we use the 95th percentile Hausdorff Distance (HD95) and Average Symmetric Surface Distance (ASSD), which are reviewed in detail by Taha and Hanbury [27].
For classification, let pi be the predicted probability for the positive class and yi∈{0 , 1} be the true label. We measure ranking quality with ROC-AUC and, for imbalanced classes, PR-AUC [28]. We assess calibration using the Brier score [29], which is the mean squared error of probabilistic forecasts, and visualize it with reliability diagrams, quantifying miscalibration with the Expected Calibration Error (ECE) [22].
To ensure full auditability and scientific reproducibility, every execution of the pipeline generates a JSON manifest file. This manifest records the software version, Git commit hash, timestamp, the selected ONNX Runtime EP, model opset version, and all computed evaluation metrics [24].
The system also provides an export module that saves segmentation masks as NIfTI files, qualitative overlays as PNG images, and all metrics in CSV/JSON formats. This functionality is managed through a comprehensive export module (see Appendix, Figure A.5). This practice aligns with best practices for reproducible computational science.
We evaluated the system on the ACDC dataset [15] for segmentation and diagnosis, and the M&Ms-2 dataset [30] for cross-domain generalization.
The SKIF-Seg module demonstrates robust performance. Table 1 presents a structure-wise comparison with a U-Net baseline. Our approach yields a significant improvement in boundary delineation, reducing the HD95 for the Left Ventricle (LV) from 7.5 mm to 5.8 mm.
The distribution of Dice scores is visualized in Figure 2, showing reduced variance for the proposed method.
Segmentation accuracy by dataset and structure (SKIF-Seg) 1.000 tn0.975 e i c if0.950 f e o c0.925 y t i r la0.900 i m is0.875 e c i D0.850 0.825
Table 2 summarizes the macro-averaged performance, highlighting a 1.67 mm reduction in HD95 on the ACDC dataset.
We compared our system against leading methods, including nnU-Net and MedNeXt (Table 3). Our system achieves a mean Dice of 0.939, matching MedNeXt and remaining highly competitive with nnU-Net, while operating within a portable ONNX framework.
To evaluate robustness, we analyzed the domain shift from ACDC to M&Ms-2 (Figure 4). The degradation in Dice scores is minimal (< 0.013), indicating excellent generalization capabilities across different scanner vendors.
The KI-GCN module demonstrates high diagnostic accuracy. Figure 5 displays the Macro ROC and PR curves, with an AUC of 0.964.
Macro ROC curve (multiclass, one-vs-rest) 1.0 Macro precision-recall curve 0.8 e t a r e0.6 v i t i s o p0.4 e u r T 0.2 0.0
ROC (AUC = 0.935) 0.0 0.2
0.4 0.6 False positive rate 0.8
1.0 (a) 1.0 0.8
The confusion matrix (Figure 6) shows strong discrimination between all five cardiac conditions.
Model trustworthiness was assessed via reliability diagrams (Figure 7). Post-hoc temperature scaling (τ =2.1) significantly improved calibration, reducing the Expected Calibration Error (ECE) to 0.03 (Table 4). Post (temp. scaling, τ =2.1)
Brier ↓ 0.08 0.07
ECE ↓ 0.04 0.03
The ablation study in Table 5 confirms that the inclusion of the graph module (KI-GCN) contributes significantly to accuracy compared to a baseline MLP. 4 5
Batch size
Finally, system throughput is analyzed in Table 6 and Figure 8. The CUDA and DirectML providers offer substantial speedups over CPU, enabling real-time clinical use.
Throughput vs. batch size by execution provider
The results of this study underscore the critical importance of a holistic systems engineering
approach to medical AI. While pure algorithmic research often prioritizes incremental gains in Dice
scores [
Our key scientific finding is the efficacy of the KI-GCN module. By explicitly modeling the heart as a graph of connected structures, we achieved a 4.9% improvement in diagnostic accuracy over a feature-based MLP baseline. This validates the hypothesis that structural knowledge is a powerful inductive bias. Furthermore, the strong calibration results (ECE of 0.03) suggest that the system’s probability outputs are trustworthy, a prerequisite for use in high-stakes medical decision-making.
However, the system is not without limitations. The current graph topology in KI-GCN is static, defined by a priori anatomical knowledge. This prevents the model from discovering novel, datadriven relationships that might exist in diverse pathologies. Additionally, while the M&Ms-2 generalization results are promising, true clinical robustness requires validation across a broader spectrum of imaging artifacts and patient demographics.
Future research will focus on two avenues: (i) developing dynamic graph learning techniques that can infer patient-specific topological connections, and (ii) conducting prospective multi-site clinical trials to validate the system’s impact on diagnostic workflow efficiency and accuracy.
In this paper, we have successfully bridged the “last-mile” gap separating high-performance AI research from tangible clinical utility. By architecting a holistic intelligent information system, we resolved the tripartite challenges of data interoperability, hardware fragmentation, and model interpretability. Our solution moves beyond isolated algorithm development to provide a unified, end-to-end pipeline that seamlessly integrates standards-compliant DICOM and NIfTI ingestion, automated privacy preservation, and hardware-agnostic inference via ONNX Runtime. The empirical validation of this framework underscores its potential to transform diagnostic workflows without disrupting existing hospital infrastructure. Specifically, the proposed SKIF-Seg module demonstrated better anatomical delineation, achieving a mean Dice Similarity Coefficient of 0.939 on the ACDC benchmark, effectively matching specialized research models within a portable container. Moreover, the integration of structured domain knowledge through the novel KI-GCN classification module yielded a diagnostic accuracy of 94.0% and, critically, a low Brier score of 0.07. These metrics establish that incorporating graph-based anatomical reasoning not only enhances predictive performance but also ensures the calibration and trustworthiness essential for highstakes medical decision-making. Consequently, this study offers a scientifically reproducible and legally auditable blueprint for deploying AI in diverse hospital environments.
Future research will focus on evolving this framework from a static deployment tool into a dynamic, continuous learning ecosystem.
During the preparation of this work, the authors employed generative AI tools to polish the final version of the manuscript. Specifically, Gemini 3 Pro (owned by Google LLC) and Grammarly (owned by Grammarly, Inc.) were utilized to improve grammar, spelling, and overall readability. After using these tools, the authors reviewed and edited the content as needed and take full responsibility for the publication’s content. systems & technologies (ICST 2024), CEUR-WS.org, Aachen, 2024, pp. 262–272. URL: https://ceur-ws.org/Vol-3790/paper23.pdf. [13] NVIDIA Corporation, CUDA C++ programming guide, 2025. URL: https://docs.nvidia.com/cuda/cuda-c-programming-guide/. [14] Microsoft Corporation, DirectML overview, 2025. URL: https://learn.microsoft.com/en-us/windows/ai/directml/overview. [15] O. Bernard, A. Lalande, C. Zotti, F. Cervenansky, X. Yang, P.-A. Heng, Others, Deep learning techniques for automatic MRI cardiac multi-structures segmentation and diagnosis: Is the problem solved?, IEEE Trans. Med. Imaging 37.11 (2018) 2514–2525. doi:10.1109/TMI.2018.2837502. [16] T. N. Kipf, M. Welling, Semi-supervised classification with graph convolutional networks, in: International conference on learning representations (ICLR), 2017, pp. 1–14. URL: https://openreview.net/pdf?id=SJU4ayYgl. [17] M. D. Alanazi, K. Kaaniche, M. Albekairi, T. M. Alanazi, G. Abbas, Graph attention neural network for advancing medical imaging by enhancing segmentation and classification, Eng. Appl. Artif. Intell. 161 (2025) 112372. doi:10.1016/j.engappai.2025.112372. [18] G. Hinton, O. Vinyals, J. Dean, Distilling the knowledge in a neural network, arXiv Prepr. arXiv (2015). doi:10.48550/arXiv.1503.02531. [19] A. Moslemi, A. Briskina, Z. Dang, J. Li, A survey on knowledge distillation: Recent advancements, Mach. Learn. With Appl. 18 (2024) 100605. doi:10.1016/j.mlwa.2024.100605. [20] O. Chaban, E. Manziuk, P. Radiuk, Method of adaptive knowledge distillation from multiteacher to student deep learning models, J. Edge Comput. 4.2 (2025) 1–20. doi:10.55056/jec.978. [21] O. Chaban, E. Manziuk, O. Markevych, S. Petrovskyi, P. Radiuk, EMTKD at the edge: An adaptive multi-teacher knowledge distillation for robust cardiac MRI classification, in: Proceedings of the 5th edge computing workshop (DOORS 2025), CEUR-WS.org, Aachen, 2025, pp. 42–47. URL: https://ceur-ws.org/Vol-3943/paper09.pdf. [22] A. Niculescu-Mizil, R. Caruana, Predicting good probabilities with supervised learning, in: Proceedings of the 22nd international conference on machine learning (ICML), 2005, pp. 625– 632. doi:10.1145/1102351.1102430. [23] M. Xiong, A. Deng, P. W. Koh, J. Wu, S. Li, J. Xu, B. Hooi, Proximity-informed calibration for deep neural networks, in: Proceedings of the 37th international conference on neural information processing systems (neurips), 2023, pp. 68511–68538. URL: https://dl.acm.org/doi/10.5555/3666122.3669118. [24] O. Chaban, E. Manziuk, P. Radiuk, IDK medical AI: An open-source framework for AI-driven medical imaging analysis, 2025. URL: https://github.com/radiukpavlo/idk-medical-ai. [25] L. R. Dice, Measures of the amount of ecologic association between species, Ecology 26.3 (1945) 297–302. doi:10.2307/1932409. [26] P. Jaccard, Étude comparative de la distribution florale dans une portion des Alpes et du Jura,
Bull. Soc. Vaudoise Sci. Nat. 37 (1901) 547–579. doi:10.5169/SEALS-266450. [27] A. A. Taha, A. Hanbury, Metrics for evaluating 3D medical image segmentation: Analysis, selection, and tool, BMC Med. Imaging 15.29 (2015) 1–28. doi:10.1186/s12880-015-0068-x. [28] T. Saito, M. Rehmsmeier, The precision-recall plot is more informative than the ROC plot when evaluating binary classifiers on imbalanced datasets, PLOS ONE 10.3 (2015) e0118432. doi:10.1371/journal.pone.0118432. [29] G. W. Brier, Verification of forecasts expressed in terms of probability, Mon. Weather Rev. 78.1 (1950) 1–3. doi:10.1175/1520-0493(1950)078%3C0001:VOFEIT%3E2.0.CO;2. [30] V. M. Campello, P. Gkontra, C. Izquierdo, C. Martín-Isla, A. Sojoudi, P. M. Full, K. MaierHein, Y. Zhang, Z. He, Others, Multi-centre, multi-vendor and multi-disease cardiac segmentation: The M&Ms challenge, IEEE Trans. Med. Imaging 40.12 (2021) 3543–3554. doi:10.1109/TMI.2021.3090082.
This appendix provides select screenshots from the graphical user interface of the IDK Medical AI system, illustrating the key stages of the end-to-end workflow.
main user interface of the IDK
Medical
AI system, providing access to data ingestion
modules (DICOM/NIfTI), analysis pipelines (Segmentation, Classification), and project management features. of DICOM series and applies privacy-preserving profiles.
Сегментація SKIF-Seg Автоматична сегментація медичних зображень з використанням нейронної мережі Вибір моделі ./models/skifseg.onnx opset=17
Validated Параметри Поріг бінаризації Morphology post-proc CRF refinement
Normalize to [
CPU Візуалізація
Browse... graph source and initiate the graph-based diagnostic classification.
Run segmentation
Inference time per slice (ms)
Research use only. Not a medical device.
Figure A.3: Interface for the SKIF-Seg segmentation module. Users can select an ONNX model and monitor the segmentation progress.
Класифікація KI-GCN Аналіз медичних зображень за допомогою графової нейронної мережі Налаштування
Результати Параметри моделі Graph source ./datasets/graph.json Features
auto Threshold 0.70
CPU Провайдер інференсу Run classification 1
2
3 Підтвердження Маски сегментації (NIfTI/DICOM) Експорт масок сегментації у форматі NIfTI або DICOM Help users to export segmentation masks (NIfTI), visual overlays, and metrics.
Метрики якості (CSV) Метрика Dice Jaccard/IoU JSON-супровід