{"id":1726,"date":"2026-03-09T08:58:45","date_gmt":"2026-03-09T08:58:45","guid":{"rendered":"https:\/\/hria.utcluj.ro\/?p=1726"},"modified":"2026-06-15T12:57:38","modified_gmt":"2026-06-15T12:57:38","slug":"arhitecturi-agentice-si-modele-de-deep-learning-in-cadrul-romanian-ai-hospital","status":"publish","type":"post","link":"https:\/\/hria.utcluj.ro\/en\/arhitecturi-agentice-si-modele-de-deep-learning-in-cadrul-romanian-ai-hospital\/","title":{"rendered":"Romanian AI Hospital \u2013 agentic architectures and deep learning models"},"content":{"rendered":"

Agentic AI: Video Segmentation for Polyp Detection (Gastro Clinic)<\/strong><\/h2>\n\n\n\n

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

Architecture and Methodology<\/strong><\/h3>\n\n\n\n
    \n
  • Backbone:<\/strong> ConvNeXt<\/strong> is used, a CNN-based architecture (ResNet) but improved with Transformer-specific design principles.<\/li>\n\n\n\n
  • Feature Extraction:<\/strong> 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).<\/li>\n\n\n\n
  • KAN Innovation:<\/strong> The architecture was enhanced by adding 3 Kolmogorov-Arnold (KAN)<\/strong> layers.<\/li>\n\n\n\n
  • KAN Layer Configuration:<\/strong> A KAN layer is defined as a matrix of activation functions and uses depth-wise<\/em> convolutions.<\/li>\n\n\n\n
  • Training:<\/strong> The ConvNeXt backbone weights were obtained through pre-training on the ImageNet-22K<\/strong> dataset.<\/li>\n<\/ul>\n\n\n
    \n
    \"\"<\/figure>\n<\/div>\n\n\n

    Performance and Validation<\/strong><\/h3>\n\n\n\n
      \n
    • Dataset:<\/strong> The evaluation was performed on SUN-SEG<\/strong>, the largest fully annotated colonoscopic sequence dataset, using 19,544 images containing polyps and their corresponding masks.<\/li>\n\n\n\n
    • Reference metrics (Dice SOTA):<\/strong> Cross-Attention-based models use PNS+ (73.7%), SALI (82.2%), and YOLO-SAM2 (90.2%) as benchmarks.<\/li>\n\n\n\n
    • Proposed Model Results:<\/strong> It outperforms PNS+ on every metric for both difficulty levels of the dataset.<\/li>\n\n\n\n
    • Computational Efficiency:<\/strong> The model maintains a processing speed above 10 FPS, reaching an average of 12.91 FPS<\/strong>.<\/li>\n<\/ul>\n\n\n\n

      Automatic Invoice Management & Resident Support System<\/strong><\/h2>\n\n\n\n

      The system integrates administrative document processing and academic support through neuro-symbolic methods and ontologies.<\/sup><\/sup>.<\/p>\n\n\n\n

      Automated Invoice Management<\/strong><\/h3>\n\n\n\n
        \n
      • Data ingestion:<\/strong> Invoices are automatically downloaded from SPV (Private Virtual Space) and processed into a knowledge graph aligned with the system ontology.<\/li>\n\n\n\n
      • Ontological Capabilities:<\/strong> The integration of product and invoice ontologies enables update suggestions based on invoiced content.<\/li>\n\n\n\n
      • GRN Automation:<\/strong> Automatic generation of Goods Receipt Notes (GRN) based on history and product characteristics from the ontology.<\/li>\n<\/ul>\n\n\n\n

        Resident Support System<\/strong><\/h3>\n\n\n\n
          \n
        • Neuro-Symbolic Approach:<\/strong> File classification and annotation use the ontology as a foundation, while the actual classification is performed using a neural approach.<\/li>\n\n\n\n
        • Query:<\/strong> The search function is implemented through SPARQL<\/strong> queries over the system ontology.<\/li>\n\n\n\n
        • Development Status:<\/strong> The complexity of the ontology is being analyzed and the appropriate reasoning engine is being selected (GraphDB, Stardog, or more expressive solutions).<\/li>\n<\/ul>\n\n\n\n

          Feedback-Driven Agentic LLM for Autonomous Research<\/strong><\/h2>\n\n\n\n

          The multi-agent architecture is designed for structured academic workflows and autonomous research.<\/sup><\/sup><\/sup><\/sup>.<\/p>\n\n\n\n

          Workflow (Pipeline)<\/strong><\/h3>\n\n\n\n

          The system uses JSON contracts<\/strong> for communication between agents and includes editorial-style revision loops<\/sup><\/sup><\/sup>:<\/p>\n\n\n\n

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

            Technical Objectives<\/strong><\/h3>\n\n\n\n
              \n
            • Reproducible Reasoning:<\/strong> Documentation of the reasoning process to ensure reproducibility.<\/li>\n\n\n\n
            • Final Output:<\/strong> Generation of a one-page summary, including the list of references and the corresponding citations for the analyzed topic.<\/li>\n<\/ul>\n\n\n\n

              <\/p>","protected":false},"excerpt":{"rendered":"

              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 SOTA (State-of-the-Art) architectures. Architecture and Methodology Performance and Validation Automatic Invoice Management & Resident Support System The system integrates administrative document processing and academic support through neuro-symbolic methods and ontologies\u2026 <\/p>","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"neve_meta_sidebar":"","neve_meta_container":"default","neve_meta_enable_content_width":"on","neve_meta_content_width":100,"neve_meta_title_alignment":"","neve_meta_author_avatar":"off","neve_post_elements_order":"[\"content\"]","neve_meta_disable_header":"off","neve_meta_disable_footer":"","neve_meta_disable_title":"","_themeisle_gutenberg_block_has_review":false,"footnotes":""},"categories":[7],"tags":[],"class_list":["post-1726","post","type-post","status-publish","format-standard","hentry","category-rezultate"],"_links":{"self":[{"href":"https:\/\/hria.utcluj.ro\/en\/wp-json\/wp\/v2\/posts\/1726","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/hria.utcluj.ro\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/hria.utcluj.ro\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/hria.utcluj.ro\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/hria.utcluj.ro\/en\/wp-json\/wp\/v2\/comments?post=1726"}],"version-history":[{"count":5,"href":"https:\/\/hria.utcluj.ro\/en\/wp-json\/wp\/v2\/posts\/1726\/revisions"}],"predecessor-version":[{"id":1865,"href":"https:\/\/hria.utcluj.ro\/en\/wp-json\/wp\/v2\/posts\/1726\/revisions\/1865"}],"wp:attachment":[{"href":"https:\/\/hria.utcluj.ro\/en\/wp-json\/wp\/v2\/media?parent=1726"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/hria.utcluj.ro\/en\/wp-json\/wp\/v2\/categories?post=1726"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/hria.utcluj.ro\/en\/wp-json\/wp\/v2\/tags?post=1726"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}