Learning Normalized Inputs for Iterative Estimation in Medical Image Segmentation (1702.05174v1)

Abstract: In this paper, we introduce a simple, yet powerful pipeline for medical image segmentation that combines Fully Convolutional Networks (FCNs) with Fully Convolutional Residual Networks (FC-ResNets). We propose and examine a design that takes particular advantage of recent advances in the understanding of both Convolutional Neural Networks as well as ResNets. Our approach focuses upon the importance of a trainable pre-processing when using FC-ResNets and we show that a low-capacity FCN model can serve as a pre-processor to normalize medical input data. In our image segmentation pipeline, we use FCNs to obtain normalized images, which are then iteratively refined by means of a FC-ResNet to generate a segmentation prediction. As in other fully convolutional approaches, our pipeline can be used off-the-shelf on different image modalities. We show that using this pipeline, we exhibit state-of-the-art performance on the challenging Electron Microscopy benchmark, when compared to other 2D methods. We improve segmentation results on CT images of liver lesions, when contrasting with standard FCN methods. Moreover, when applying our 2D pipeline on a challenging 3D MRI prostate segmentation challenge we reach results that are competitive even when compared to 3D methods. The obtained results illustrate the strong potential and versatility of the pipeline by achieving highly accurate results on multi-modality images from different anatomical regions and organs.

• The paper introduces a two-stage pipeline that uses an FCN for input normalization and an FC-ResNet for iterative segmentation refinement.
• Experimental results on electron microscopy, CT liver lesions, and MRI prostate images demonstrate improved segmentation accuracy and robust performance.
• The approach minimizes pre-processing complexity by learning data normalization, enabling versatile applications across a range of medical imaging modalities.

Learning Normalized Inputs for Iterative Estimation in Medical Image Segmentation

The paper, "Learning Normalized Inputs for Iterative Estimation in Medical Image Segmentation," presents an enhanced pipeline for medical image segmentation by integrating Fully Convolutional Networks (FCNs) with Fully Convolutional Residual Networks (FC-ResNets). This approach offers a novel perspective on leveraging recent developments in understanding Convolutional Neural Networks (CNNs) and ResNets to achieve state-of-the-art performance in medical image segmentation tasks.

Overview of the Proposed Pipeline

The proposed pipeline utilizes a low-capacity FCN model as a pre-processor to normalize input images, preparing them for the FC-ResNet, which iteratively refines these inputs to derive accurate segmentation predictions. This dual-model setup benefits from the strengths of both applied architectures: the FCN efficiently preprocesses input images, while the FC-ResNet, with its layered depths, performs as a powerful segmentor through iterative estimations. The segmentation pipeline is notably adaptable across different imaging modalities, including 2D electron microscopy, CT scans for liver lesions, and 3D MRI prostate images, demonstrating its versatility and robustness.

Experimental Results and Observations

• Electron Microscopy Benchmark: The pipeline achieved notable performance on the ISBI 2012 EM dataset, exhibiting superior results among 2D methods. The FCN pre-processor's role in normalizing the data positively influenced the FC-ResNet's performance in generating precise segmentations.
• CT Liver Lesion Segmentation: The pipeline outperformed conventional FCN and existing FC-ResNet methods in segmenting liver lesions from CT scans. The normalization strategy was particularly effective in addressing the high variability inherent in CT image intensities.
• MRI Prostate Segmentation: Applied without traditional pre-processing, the proposed pipeline delivered competitive results in the PROMISE12 challenge, even surpassing some 3D methods. This highlights the FCN’s ability to learn effective normalization across diverse input magnitudes and resolutions.

Theoretical and Practical Implications

The integration of FCNs and FC-ResNets reflects a two-fold approach: simplifying data pre-processing through trainable, data-driven transformation and utilizing the depth and connectivity of ResNets for complex, iterative refinement tasks. The findings underscore the potential for creating deeper FC-ResNets using learned input normalization, which could lead to more efficient segmentations in real-world clinical scenarios where data pre-processing can be a bottleneck.

Future Directions

Future work may focus on exploring other architectural variants for the pre-processing step, potentially expanding the system to incorporate 3D FCNs, thus exploiting volumetric data directly. Further research could also investigate the theoretical underpinnings of iterative refinement in fully convolutional architectures, particularly in their application across various imaging modalities and medical conditions.

In summation, this paper demonstrates a successful incorporation of sophisticated neural network architectures to yield a flexible and powerful medical image segmentation tool, showing promise for extensive clinical application across different imaging modalities and anatomical regions.