Skip to content

Romanian AI Hospital – agentic architectures and deep learning models

    Agentic AI: Video Segmentation for Polyp Detection (Gastro Clinic)

    The implementation aims to optimize polyp detection through real-time video segmentation, using a colonoscopic agent based on state-of-the-art (SOTA) architectures.

    Architecture and Methodology

    • Backbone: ConvNeXt is used, a CNN-based architecture (ResNet) but improved with Transformer-specific design principles.
    • Feature Extraction: The model is structured into 4 stages; the last two provide low-level features (H/8 x W/8) and high-level features (H/16 x W/16).
    • KAN Innovation: The architecture was enhanced by adding 3 Kolmogorov-Arnold (KAN) layers.
    • KAN Layer Configuration: A KAN layer is defined as a matrix of activation functions and uses depth-wise convolutions.
    • Training: The ConvNeXt backbone weights were obtained through pre-training on the ImageNet-22K dataset.

    Performance and Validation

    • Dataset: The evaluation was performed on SUN-SEG, the largest fully annotated colonoscopic sequence dataset, using 19,544 images containing polyps and their corresponding masks.
    • Reference metrics (Dice SOTA): Cross-Attention-based models use PNS+ (73.7%), SALI (82.2%), and YOLO-SAM2 (90.2%) as benchmarks.
    • Proposed Model Results: It outperforms PNS+ on every metric for both difficulty levels of the dataset.
    • Computational Efficiency: The model maintains a processing speed above 10 FPS, reaching an average of 12.91 FPS.

    Automatic Invoice Management & Resident Support System

    The system integrates administrative document processing and academic support through neuro-symbolic methods and ontologies..

    Automated Invoice Management

    • Data ingestion: Invoices are automatically downloaded from SPV (Private Virtual Space) and processed into a knowledge graph aligned with the system ontology.
    • Ontological Capabilities: The integration of product and invoice ontologies enables update suggestions based on invoiced content.
    • GRN Automation: Automatic generation of Goods Receipt Notes (GRN) based on history and product characteristics from the ontology.

    Resident Support System

    • Neuro-Symbolic Approach: File classification and annotation use the ontology as a foundation, while the actual classification is performed using a neural approach.
    • Query: The search function is implemented through SPARQL queries over the system ontology.
    • Development Status: The complexity of the ontology is being analyzed and the appropriate reasoning engine is being selected (GraphDB, Stardog, or more expressive solutions).

    Feedback-Driven Agentic LLM for Autonomous Research

    The multi-agent architecture is designed for structured academic workflows and autonomous research..

    Workflow (Pipeline)

    The system uses JSON contracts for communication between agents and includes editorial-style revision loops:

    1. Researcher: Receives the research task and collects the data.
    2. Writer: Produces the content based on the research.
    3. Fact Checker: Performs claim validation and ensures claim-level evidence traceability.
    4. Editor: Refines the final output; may request a redo if quality criteria are not met.

    Technical Objectives

    • Reproducible Reasoning: Documentation of the reasoning process to ensure reproducibility.
    • Final Output: Generation of a one-page summary, including the list of references and the corresponding citations for the analyzed topic.