`Project & Excite' Modules for Segmentation of Volumetric Medical Scans (1906.04649v2)

Abstract: Fully Convolutional Neural Networks (F-CNNs) achieve state-of-the-art performance for image segmentation in medical imaging. Recently, squeeze and excitation (SE) modules and variations thereof have been introduced to recalibrate feature maps channel- and spatial-wise, which can boost performance while only minimally increasing model complexity. So far, the development of SE has focused on 2D images. In this paper, we propose Project & Excite' (PE) modules that base upon the ideas of SE and extend them to operating on 3D volumetric images.Project & Excite' does not perform global average pooling, but squeezes feature maps along different slices of a tensor separately to retain more spatial information that is subsequently used in the excitation step. We demonstrate that PE modules can be easily integrated in 3D U-Net, boosting performance by 5% Dice points, while only increasing the model complexity by 2%. We evaluate the PE module on two challenging tasks, whole-brain segmentation of MRI scans and whole-body segmentation of CT scans. Code: https://github.com/ai-med/squeeze_and_excitation

• The paper introduces the Project Excite (PE) module, an innovative computational block for 3D fully convolutional neural networks designed to improve the accuracy of segmenting volumetric medical scans.
• The PE module enhances performance by recalibrating feature maps across slice projections, preserving vital spatial information through tensor slicing unlike traditional global pooling methods, and adding only 2% model complexity.
• Experimental results show PE modules integrated into a 3D U-Net yielded a 5% increase in Dice score for whole-brain MRI and whole-body CT segmentation compared to baselines, particularly benefiting segmentation of small structures with limited data.

Analysis of `Project Excite' Modules for Segmentation of Volumetric Medical Scans

The paper introduces an innovative computational module, the `Project Excite' (PE) module, tailored for fully convolutional neural networks (F-CNNs) to enhance their performance in segmenting 3D volumetric medical scans. Drawing upon the principles of the established squeeze and excitation (SE) block concept, the authors extend its applicability from 2D to 3D data by retaining more spatial information, which is vital for precise anatomical localization in medical imaging tasks.

Methodology

The `Project Excite' (PE) module distinguishes itself by omitting global average pooling, a staple in traditional SE modules, and instead utilizes tensor slicing to preserve spatial information across different tensor planes. The PE module recalibrates feature maps in 3D F-CNNs, like the 3D U-Net, enhancing segmentation accuracy with only a 2% increase in model complexity. This recalibration is achieved by learning dependencies across slice projections of the feature maps, in stark contrast to mere channel dependencies.

The authors meticulously evaluate their approach through two significant segmentation challenges: whole-brain segmentation in MRI scans and whole-body segmentation in CT scans. Implementing PE in these tasks yielded a notable enhancement of 5% in Dice score compared to conventional 3D U-Nets, highlighting the module's efficacy.

Experimental Setup and Results

In terms of experimental setup, the authors employ a 3D U-Net architecture with various modifications to effectively utilize whole 3D scans during training. To preserve manageable model complexity, the number of channels per layer is reduced. The training utilizes instance normalization and an array of data augmentation techniques to better adapt to the limited datasets available, a common scenario in medical imaging.

The results presented are substantiated by a granular investigation into the placement of PE blocks within the network, showing that their strategic integration across encoder, decoder, and bottleneck layers yields optimal performance enhancements. When compared against both the baseline 3D U-Net and the traditional 3D channel SE (cSE) blocks, the PE module not only boosts performance significantly in segmenting small and challenging anatomical structures, but also exhibits superior efficacy in the context of limited dataset training, outperforming added convolutional layers or alternative recalibration modules.

Implications and Future Directions

The PE module's capacity to enhance segmentation with marginal computational overhead makes it an attractive proposition for integrating into existing 3D F-CNN frameworks, particularly in medical imaging contexts constrained by data scarcity and computational resources. This paper's methodological advancements could inspire further explorations into refined recalibration modules and scalable deployment strategies tailored for specific medical imaging applications.

Future research may investigate refinements to further optimize the balance between model complexity and performance gain, delve into applications across other domains of volumetric data analysis, or explore combinations with other architectural innovations to explore synergies with complementary technologies in computer-aided diagnostics. The `Project Excite' module sets a potential precedent for future work seeking to leverage detailed spatial feature interdependencies within volumetric data for enhanced computational imaging results.