AI Platform for Precision Medicine in Brain Oncology
A 24-month research project developing a comprehensive platform for analyzing clinical data from brain tumor patients through Deep Learning algorithms, enabling personalized diagnosis and automated classification.
The Challenge
Brain tumor diagnosis requires integrating information from multiple sources: histological images, radiological images, clinical data, and omics data (transcriptomics, epigenomics, genomics). Medical professionals need tools that unify these heterogeneous data sources and assist in clinical decision-making, especially when dealing with limited datasets and strict privacy requirements.
My Role
As AI researcher and platform developer, I led the complete technical development:
- Designed and deployed the full clinical analysis platform
- Researched and developed Deep Learning models specialized in histological and radiological images
- Implemented distributed training pipelines using PyTorch Lightning
- Designed data augmentation and transfer learning strategies for small medical datasets
- Developed explainability techniques for clinical decision support
- Scientifically validated algorithms with metrics adapted to the medical context
Technical Approach
Multimodal analysis:
- Integration of clinical data, medical images, and omics data into a unified pipeline
- Specialized models for each data modality
- Processing of transcriptomics, epigenomics, and oncological genomics
Backend and ML:
- PyTorch + Lightning as the main Deep Learning framework
- FastAPI as a high-performance REST API for serving predictions
- PostgreSQL for structured clinical data storage
- Distributed training pipeline with automated validation and evaluation
Frontend:
- React for an intuitive clinical interface aimed at medical professionals
- Visualization of analysis results and predictions
Technical Challenges
- Multimodality: Effective integration of heterogeneous data (images, text, omics) into a unified model
- Limited data: Data augmentation and transfer learning strategies to compensate for small datasets
- Medical privacy: Compliance with healthcare data protection regulations (GDPR, Spanish healthcare regulation)
- Explainability: Implementation of interpretability techniques so clinicians can trust model decisions
- Clinical validation: Evaluation metrics and protocols adapted to real-world medical context
Results
The platform enables doctors and researchers to:
- Accelerate brain tumor diagnosis through automated classification
- Identify new therapeutic biomarkers from multi-omics analysis
- Personalize treatments based on the patient’s molecular profile
- Facilitate collaborative research in precision medicine
Funding
Project funded by the Valencian Innovation Agency (AVI) under file INNEST/2022/159, in collaboration with ISABIAL, FISABIO, AIJU, and The New Horizons.