AMD Life Science 26.08 ships complete open-source medical imaging stack on ROCm 10.0

AMD has released Life Science 26.08, a vertically integrated medical imaging stack built on ROCm 10.0 that accelerates the full pipeline from disk read to clinician annotation on AMD Instinct GPUs. The release combines four components: hipCIM 26.06 for GPU-accelerated whole-slide and 3D medical image I/O with cuCIM API compatibility; MONAI Core 1.6.0 for ROCm-validated inference; MONAI Model Zoo 26.08 with five validated, optimized model bundles; and MONAILabel 0.8.5 for interactive active-learning annotation. The stack achieves approximately 10x faster whole-slide image patch reads versus OpenSlide, 4.2x faster MONAI data loading, and 1.6-3.8x faster end-to-end pathology inference, while maintaining API compatibility so existing MONAI and cuCIM code requires no changes.

The release is structured as a capstone to five companion blog posts, each detailing a layer of the stack. hipCIM notably expands format coverage from 0.8% to 90.3% for OME-TIFF and adds 3D NIfTI and DICOM support, plus GPU decode of JPEG-compressed Aperio SVS slides. MONAI Model Zoo optimizations are delivered as non-invasive inference_rocm.json/yaml overlays that leave upstream bundles and other runtimes untouched. MONAILabel brings 131x fp32 inference acceleration for interactive annotation workflows.

Stack architecture and ROCm integration

• hipCIM 26.06 I/O layer: GPU-accelerated whole-slide image (WSI) and 3D medical image reading via the from cucim import CuImage API, achieving 9,661-9,930 patches/second on MI355X with ROCm 10.0 versus ~950 p/s for OpenSlide GPU path; CPU path reaches 7-11x speedup on the same node.
• Format expansion and GPU decode: OME-TIFF coverage increases from 0.8% to 90.3%; adds 3D NIfTI and DICOM reading through identical cucim.CuImage API; introduces GPU decode of real JPEG-compressed Aperio SVS slides from Leica Aperio scanners.
• MONAI Core 1.6.0 integration: hipCIM serves as MONAI's WSIReader backend, yielding ~4.2x faster data loading (~2,750 p/s vs ~660 p/s) and 1.6-3.8x faster end-to-end pathology inference across three slides at num_workers=8, with stable ~0.13 s/batch read cost versus variable 0.21-0.51 s/batch for OpenSlide.
• MONAI Model Zoo validation: Five bundles validated for inference on MI355X (gfx950) with inference times of 3.60-3,287 ms, all passing Dice metrics; optimizations ship as inference_rocm.json/inference_rocm.yaml overlays without modifying upstream bundle configurations.
• MONAILabel 0.8.5 annotation server: Interactive active-learning annotation with 131x fp32 inference acceleration (median of 4 spleen CT volumes), completing the pipeline from image ingestion through model inference to clinician review and refinement.

Why it matters

This release strengthens AMD's position in open-source AI healthcare tooling by delivering a ROCm-native alternative to NVIDIA's dominant CUDA-based medical imaging ecosystem, using the same PyTorch and MONAI APIs that researchers already employ. For Linux-based HPC and research environments, the preservation of existing cucim and MONAI import paths means migration friction is minimized, while the overlay-based optimization approach for model bundles avoids the fork-and-patch maintenance burden that often fragments open-source AI tooling. The explicit validation against Dice metrics and publication of system-under-test methodology in companion posts provides the reproducibility baseline that clinical and regulatory contexts require, potentially accelerating adoption in regulated healthcare deployments where AMD Instinct hardware is increasingly competitive.