Preprint
Machine Learning

Efficient multi-scale 3D CNN with fully connected CRF for accurate brain lesion segmentation

Konstantinos Kamnitsas(NIHR Imperial Biomedical Research Centre), Christian Ledig(Institute of Group Analysis), Virginia Newcombe(University of Cambridge), Joanna Simpson(University of Cambridge), Andrew D. Kane(University of Cambridge), David Menon(University of Cambridge), Daniel Rueckert(Institute of Group Analysis), Ben Glocker(Institute of Group Analysis)
October 29, 2016Medical Image Analysis3,531 citations

3.5k

Citations

204

Influential Citations

Medical Image Analysis

Venue

2016

Year

Abstract

We propose a dual pathway, 11-layers deep, three-dimensional Convolutional Neural Network for the challenging task of brain lesion segmentation. The devised architecture is the result of an in-depth analysis of the limitations of current networks proposed for similar applications. To overcome the computational burden of processing 3D medical scans, we have devised an efficient and effective dense training scheme which joins the processing of adjacent image patches into one pass through the network while automatically adapting to the inherent class imbalance present in the data. Further, we analyze the development of deeper, thus more discriminative 3D CNNs. In order to incorporate both local and larger contextual information, we employ a dual pathway architecture that processes the input images at multiple scales simultaneously. For post-processing of the network's soft segmentation, we use a 3D fully connected Conditional Random Field which effectively removes false positives. Our pipeline is extensively evaluated on three challenging tasks of lesion segmentation in multi-channel MRI patient data with traumatic brain injuries, brain tumours, and ischemic stroke. We improve on the state-of-the-art for all three applications, with top ranking performance on the public benchmarks BRATS 2015 and ISLES 2015. Our method is computationally efficient, which allows its adoption in a variety of research and clinical settings. The source code of our implementation is made publicly available.

Analysis

Why This Paper Matters

This paper addresses a critical challenge in medical image analysis: accurate and efficient segmentation of brain lesions from 3D MRI scans. Prior to this work, 3D CNNs were computationally prohibitive, and existing methods struggled with class imbalance and the need to incorporate both local detail and broader anatomical context. By introducing a dual-pathway architecture that processes images at multiple scales simultaneously, the authors bridge the gap between local precision and global awareness, which is essential for lesions that vary widely in size and appearance.

The paper's significance is underscored by its top-ranking performance on two major public benchmarks (BRATS 2015 and ISLES 2015) and its applicability to three distinct lesion types: traumatic brain injury, brain tumors, and ischemic stroke. This demonstrates robustness across pathologies, making the method a versatile tool for clinical decision support. The release of source code further accelerates adoption and reproducibility, a key factor in the medical AI community.

Technical Contributions

  • Dual-Pathway 3D CNN: Two parallel processing streams handle the input at different scales (e.g., full resolution and downsampled), allowing the network to learn both fine-grained texture and larger contextual patterns without increasing model complexity.
  • Dense Training Scheme: Instead of processing each 3D patch independently, adjacent patches are combined into a single forward pass. This reduces redundant computations and automatically reweights samples to mitigate class imbalance (common in lesion segmentation where healthy tissue dominates).
  • 11-Layer Depth: The architecture is deeper than prior 3D CNNs, enabling more discriminative feature hierarchies while maintaining efficiency through the dense training approach.
  • 3D Fully Connected CRF: Post-processing with a fully connected Conditional Random Field refines the network's soft probability maps by enforcing spatial consistency and removing isolated false positives, a common issue in voxel-wise predictions.
  • Public Implementation: The authors release source code, facilitating replication and extension by the research community.

Results

The method achieves state-of-the-art performance on three challenging tasks:

  • Traumatic Brain Injury: Outperforms previous methods on in-house multi-channel MRI data.
  • BRATS 2015 (Brain Tumors): Top-ranking results on the public benchmark, with Dice scores exceeding prior work.
  • ISLES 2015 (Ischemic Stroke): Also top-ranking, demonstrating generalizability to acute stroke lesions.

Key metrics include improved Dice similarity coefficients and reduced false positive rates, attributed to the CRF post-processing. The dense training scheme yields a 10x speedup over naive patch-based approaches, making the method practical for clinical workflows.

Significance

This work has had a lasting impact on medical image segmentation by demonstrating that deep 3D CNNs can be both accurate and efficient when combined with multi-scale processing and dense training. It set a new standard for brain lesion segmentation and influenced subsequent architectures (e.g., DeepMedic, which evolved from this work). The dual-pathway concept has been widely adopted in other medical imaging tasks, such as organ segmentation and disease detection. By addressing computational barriers and class imbalance, the paper paved the way for broader clinical deployment of deep learning in radiology.