Hospital and Healthcare HPC

The need for computing infrastructure in the healthcare sector has moved beyond the clinic. Radiology departments process hundreds of high-resolution imaging sets daily; genomics laboratories work with whole exome or whole genome sequencing data; research units train machine learning models on patient cohorts. All of these workloads require both high computing capacity and infrastructure that provides legal assurance for patient data.

In Turkey, personal health data falls under the “special category personal data” classification within KVKK. This makes it mandatory for such data to be processed domestically and in an auditable environment. Mevasis offers healthcare organizations Turkey-located, KVKK-compliant HPC infrastructure and managed service models.

Why Is HPC Necessary in Healthcare?

The acceleration in medical imaging and clinical data analysis is pushing the limits of traditional server infrastructure. Three concrete examples:

Pathology imaging: File size per digital pathology slide ranges from 2–15 GB. Running deep learning-based cancer detection models requires GPU acceleration; CPU-based processing of a single slide can take hours.
Radiomics: Extracting 500–1,000 quantitative features from CT or MRI series and training models correlating these features with clinical outcomes exceeds the capacity of standard workstations.
Clinical genomics: Obtaining a patient’s tumor mutation profile via whole exome sequencing (WES) analysis requires a pipeline that takes 24–48 hours on standard servers. This time drops to 1–2 hours with NVIDIA Parabricks.

Typical Workloads

Medical Imaging and Artificial Intelligence

Image-based diagnostic support systems require large datasets and powerful GPUs during training, while low latency is the priority during inference.

Commonly used frameworks and tools:

TensorFlow / PyTorch: Training radiology, pathology, dermatology models
MONAI (Medical Open Network for AI): Optimized PyTorch extension for medical imaging; used in segmentation, classification, and detection workloads
ITK / SimpleITK: CT and MRI image preprocessing, registration, and segmentation
3D Slicer: Image analysis and clinical research integration
DICOM tools (dcm2niix, pydicom): Converting DICOM files to analysis-ready formats

Example: Lung nodule detection model training

Hardware	100 epoch training time
16 CPU cores	~96 hours
Single NVIDIA A100 80 GB	~8 hours
4× NVIDIA A100 (multi-GPU)	~2.5 hours

Genomics and Bioinformatics

Clinical genomics is increasingly becoming routine for diagnosis and treatment planning. Tumor genome analysis, rare disease diagnosis, and pharmacogenomic tests are the primary application areas.

Tool	Function	HPC Requirement
BWA-MEM2	Short-read alignment	Multi-core CPU, fast I/O
GATK 4	Germline variant calling, somatic variant analysis	High RAM (64–256 GB/sample), parallel I/O
NVIDIA Parabricks	GATK pipeline GPU acceleration	NVIDIA GPU; 50× faster than GATK
DeepVariant	Deep learning-based variant calling	GPU (A100/H100 recommended)
STAR / HISAT2	RNA-seq alignment (gene expression analysis)	Multi-core CPU, large memory
ANNOVAR / VEP	Variant annotation	CPU-intensive, database I/O
Cromwell / Snakemake	Pipeline orchestration	Low capacity; workflow management

Clinical WES example (Whole Exome Sequencing):

Raw data: 10–15 GB FASTQ
CPU-based GATK Best Practices: 8–16 hours
Parabricks GPU pipeline: 30–45 minutes
Final output: VCF file + filtered variant list for clinical report

Molecular Simulation and Drug Research

Research centers and advanced medical faculties within hospital systems use computational methods to understand the mechanisms of action of experimental treatments.

GROMACS / NAMD: Protein-drug interaction molecular dynamics simulations; 10–30× acceleration with GPU
AutoDock Vina / GNINA: Drug candidate screening via molecular docking
AlphaFold 2 / AlphaFold 3: Target protein structure prediction; NVIDIA GPU required
Rosetta: Protein design and protein-protein interaction analysis

Clinical Data Analytics and Machine Learning

Large-scale machine learning studies conducted on patient records, laboratory results, and clinical notes:

Intensive care mortality prediction models (MIMIC dataset-compatible frameworks)
Time series analysis on electronic health records (EHR)
Clinical note coding via natural language processing (ICD-10 automation)
Federated learning infrastructure for multi-center cohort studies

KVKK Compliance and Data Security

In Turkey, personal health data processing falls under the special category data class under Article 6 of KVKK. Measures required for this class include technical and administrative controls, data processing records, and compliance with Personal Data Protection Board regulations.

Using overseas cloud infrastructure means transferring personal health data abroad. This transfer is subject to additional conditions under KVKK Article 9 (explicit consent or adequate protection decision) and carries serious legal risk in practical application.

Core guarantees of Mevasis healthcare infrastructure:

Physically located server infrastructure within Turkey’s borders
Network isolation: vLAN segmentation, closed network option for cluster nodes processing patient data
Access control: LDAP/Active Directory integration, role-based authorization (RBAC)
Audit logging: storage of user access, data transfer, and job queue logs
Storage encryption: AES-256 for data at rest, TLS 1.3 during transfer
Physical security: locked cabinets, entry logging, authorized personnel access only

Reference Configuration: Healthcare Research Center HPC

Healthcare Research Center — Reference HPC Configuration

Login / Transfer Node (2 units, load-balanced)
├── GPU Compute Nodes (4–8 units)
│   └── 2× Intel Xeon Gold 6438M + 4× NVIDIA A100 80 GB SXM
│       256 GB DDR5 system memory
│       (Image analysis, Parabricks, GROMACS, AlphaFold)
│
├── CPU Compute Nodes (8–16 units)
│   └── 2× AMD EPYC 9354 (32-core), 512 GB DDR5
│       (GATK, STAR, ANNOVAR, EHR analytics)
│
├── High-Memory Node (2 units)
│   └── 2× AMD EPYC 9654, 1.5 TB DDR5
│       (Large cohort analysis, Gaussian, large-scale variant annotation)
│
└── Storage Layer
    ├── NVMe scratch: BeeGFS, 4 GB/s+ bandwidth
    ├── Capacity storage: ZFS-based, redundant, encrypted
    └── Archive: S3-compatible object storage, lifecycle policies

Network: InfiniBand HDR100 (MPI and GPU-GPU traffic)
Job Scheduler: SLURM + project-specific partition structure
Security: Isolated network segment, encrypted storage, audit log

Mevasis Healthcare HPC Services

Mevasis offers specialized HPC services for hospital research centers, medical faculties, genomics laboratories, and health technology companies:

On-Site HPC Installation: The strongest option for KVKK compliance. Servers are located in your hospital’s or research center’s data center. MONAI, Parabricks, GATK, and MD software installation included.
Managed HPC Services: Isolated segment on Turkey-located Mevasis infrastructure; maintenance, updates, and security monitoring carried out by our team.
GPU Cluster Rental: Short-term, high-capacity GPU access for research projects or clinical AI model development periods.
HPC Consulting: Analysis of your current workload profile, hardware sizing, SLURM configuration, and software environment setup.

Frequently Asked Questions

If we will be processing patient data, do we absolutely need an on-premise installation? Not mandatory, but it provides the strongest compliance assurance. KVKK compliance can also be achieved in Mevasis’s Turkey-located managed hosting model with appropriate contractual and technical measures. The choice should be shaped by institutional policy and the sensitivity level of the data.

How is DICOM data processed on HPC infrastructure? DICOM files can be transferred directly to HPC storage. dcm2niix or pydicom is used for preprocessing; ITK/SimpleITK or MONAI is used for analysis. DICOM segmentation and anonymization steps can be incorporated into the pipeline during data transfer.

Is the 50× speed difference between Parabricks and GATK real? According to benchmark data published by NVIDIA and user experiences, 30–50× acceleration is a typical value range. Actual speedup varies by sample count, read depth, and GPU model used. Parabricks produces output consistent with GATK’s validated results.

Can federated learning infrastructure be set up? Yes. Federated learning configuration allowing each hospital to keep its data on its own infrastructure while sharing model weights is within Mevasis consulting scope. This approach both supports KVKK compliance and enables multi-institutional research collaboration.

Is integration with our hospital’s existing IT infrastructure possible? Yes. Active Directory integration, secure connectivity compatible with existing network architecture, and integration of the job queue with HIS/PACS systems can be included in the project scope.