Abstract
The DeepHealth Toolkit is a comprehensive, open-source European framework designed to bring high-performance deep learning and computer vision to the healthcare sector. By bridging the gap between data science and complex heterogeneous computing (HPC), the toolkit allows researchers to run sophisticated AI pipelines across GPUs, FPGAs, and cloud clusters without needing to be systems engineering experts. This paper details the architecture, key innovations, and performance benchmarks of the toolkit.
Introduction: Addressing the Technological Gap
While Artificial Intelligence has matured significantly, its adoption in the healthcare sector remains limited by a "technological gap." Medical datasets are massive and complex, requiring High-Performance Computing (HPC) for training. However, most data scientists are not experts in parallel programming or hardware-specific optimizations for heterogeneous architectures, such as clusters combining different brands of GPUs and FPGAs. Existing frameworks often require significant manual tuning to run efficiently on supercomputers or hybrid clouds. The DeepHealth project aims to provide a "made in Europe" solution that offers high-level abstractions, allowing researchers to focus on medical problems rather than underlying infrastructure.
The DeepHealth Toolkit Architecture
The toolkit is built on four primary pillars designed to provide a hardware-agnostic environment for medical researchers:
- European Computer Vision Library (ECVL): Specifically designed for medical imaging, supporting standard formats like DICOM and NIfTI. It includes a Hardware Abstraction Layer (HAL) to ensure code runs seamlessly across different devices.
- European Distributed Deep Learning Library (EDDL): The core engine supporting modern neural network topologies and using the ONNX format for model interoperability. It handles tensor operations in a hardware-agnostic way.
- Orchestration & Distribution: Integration with COMPSs (a task-based programming model) and StreamFlow (a workflow management system) allows for automatic distribution of training batches across supercomputer nodes or hybrid clouds.
- Front-end/Back-end: A RESTful API and a web-based GUI provide a "no-code/low-code" environment for model design, training, and monitoring.
Key Innovations and Insights
A central feature of the toolkit is the "Computing Service" Abstraction. Developers can write a model once and switch execution from a local CPU to a GPU cluster or an FPGA card by changing a single parameter. Unlike mainstream frameworks, DeepHealth adopts an "FPGA-First" Mentality, optimizing for these devices due to their energy efficiency and parallelism in clinical inference. Furthermore, the toolkit supports Hybrid Cloud Orchestration, enabling heavy training on supercomputers like Marenostrum followed by deployment to local Kubernetes-managed clouds for real-time inference.
Performance Evaluation
The EDDL was benchmarked against industry standards TensorFlow and PyTorch using architectures such as VGG16/19 and ResNet18/34/50 on the Cifar10 dataset.
The results demonstrate that EDDL offers competitive performance. For instance, on ResNet18 (GPU) without Batch Normalization, EDDL completed an epoch in 36 seconds, significantly faster than PyTorch’s 59 seconds. Across all models, test accuracy remained comparable to established frameworks, proving that high-level abstractions do not compromise model quality. The toolkit has been successfully validated on major European infrastructures, including the Marenostrum supercomputer (3,456 nodes) and the MANGO cluster.
Limitations and Future Work
Current benchmarks indicate a performance overhead when using Batch Normalization (BN) on larger topologies like VGG16, which is a priority for future optimization. The authors also plan to integrate the NVIDIA cuDNN library to further enhance performance on NVIDIA hardware. Additionally, the "model zoo" in the front-end will be expanded to include specialized architectures like U-Net for medical image segmentation.
Summary updated