PROJECTS & RESEARCH

Overview

BioAnalytiX is developing an AI assistant for medical image analysis, focusing on radiology workflows. The platform automates the entire pipeline from DICOM ingest through preprocessing, model training, and clinical validation.

Key Features

• Automated DICOM Processing: Built robust pipelines for ingesting and preprocessing medical imaging data in DICOM format
• CT Analysis Models: Developed and trained baseline models for CT scan analysis with focus on diagnostic accuracy
• Experiment Tracking: Implemented comprehensive MLOps infrastructure for versioning, monitoring, and comparing model iterations
• Clinical Validation: Leading iterative feedback loops with radiologists to refine workflows and ensure clinical relevance

Technical Stack

Partnerships

Established strategic collaborations with the General Hospital of Larissa and the University of Thessaly for research validation and clinical pilot programs.

Impact

Working towards MVP deployment in radiology departments to assist clinicians with faster, more accurate diagnostic support while maintaining full clinical oversight.

Overview

Advanced computer vision research project focused on automated classification of myopia severity from fundus imagery using state-of-the-art vision transformers and explainability techniques.

🤝 Collaboration: This research was conducted in collaboration with ACTA Lab at the University of Thessaly, leveraging their expertise in medical imaging and computer vision research.

Methodology

• Data & Validation: Conducted rigorous 10-fold cross-validation to ensure robust model performance across diverse patient populations
• Classification Task: Three-class classification problem distinguishing between normal, high myopia, and pathological myopia
• Model Architecture: Leveraged ViT-B/14 with DINOv2 pre-training, benchmarked against traditional CNN baselines (ResNet, EfficientNet)
• Explainability: Applied attention-rollout techniques to visualize which regions of the fundus the model focuses on for predictions

Key Findings

The system achieved 97% accuracy in classifying myopia severity. Vision transformers with DINOv2 pre-training significantly outperformed CNN baselines. Attention visualizations revealed the model correctly focused on clinically relevant features such as optic disc morphology and peripheral retinal changes.

Technical Stack

Clinical Relevance

This work contributes to automated screening systems that can help ophthalmologists identify high-risk patients earlier, potentially preventing vision loss from pathological myopia progression.

Overview

Comprehensive research thesis supervised by Prof. Fotios Kokkoras, exploring advanced neural network architectures and optimization strategies to improve diagnostic accuracy across multiple medical imaging modalities.

Research Focus

• Architecture Exploration: Systematic evaluation of CNNs, Vision Transformers, and hybrid architectures for medical image classification tasks
• Optimization Strategies: Investigated advanced training techniques including transfer learning, data augmentation, and domain adaptation
• Multi-Modal Learning: Explored methods to leverage multiple imaging modalities (CT, MRI, X-ray) for improved diagnostic performance
• Clinical Validation: Validated models against clinician annotations to ensure alignment with diagnostic standards

Key Contributions

Developed a framework for comparing model architectures across diverse imaging tasks, identified optimal pre-training strategies for medical domains, and demonstrated the importance of explainability for clinical adoption.

Technical Stack

Supervisor

Prof. Fotios Kokkoras
University of Thessaly, Department of Digital Systems

Impact

This foundational research directly informed the technical approach taken at BioAnalytiX and contributed to the broader understanding of how to effectively apply modern AI techniques to medical imaging challenges.