[論文レビュー] DLTK: State of the Art Reference Implementations for Deep Learning on Medical Images

Nick Pawlowski, Sofia Ira Ktena+5 more|ArXiv.org|Nov 18, 2017

COVID-19 diagnosis using AI参考文献 8被引用数 72

ひとこと要約

DLTKは、医用画像解析の一般的な深層学習アーキテクチャの検証済みで高性能なリファレンス実装をTensorFlow上に提供し、公開MICCAIデータセットで最先端のセグメンテーション結果を達成しています。本論文は、医用画像深層学習のベースラインを迅速に実験可能にするプラグアンドプレイ、API優先のアプローチを強調します。

ABSTRACT

We present DLTK, a toolkit providing baseline implementations for efficient experimentation with deep learning methods on biomedical images. It builds on top of TensorFlow and its high modularity and easy-to-use examples allow for a low-threshold access to state-of-the-art implementations for typical medical imaging problems. A comparison of DLTK's reference implementations of popular network architectures for image segmentation demonstrates new top performance on the publicly available challenge data "Multi-Atlas Labeling Beyond the Cranial Vault". The average test Dice similarity coefficient of $81.5$ exceeds the previously best performing CNN ($75.7$) and the accuracy of the challenge winning method ($79.0$).

研究の動機と目的

• Provide baseline, validated deep learning components for medical image analysis (data reading/preprocessing, model definitions, training strategies, deployment).
• Enable low-threshold access to state-of-the-art architectures (e.g., U-Net, FCN) and common losses for medical imaging tasks.
• Demonstrate competitive performance on a public MICCAI challenge dataset to encourage rapid experimentation and deployment.

提案手法

• Adopts a plug-and-play structure on TensorFlow to separate data handling, model architecture, and training strategies.
• Implements reference FCN and U-Net architectures with residual units for segmentation.
• Explores data reading strategies (random vs. class-balanced patch sampling) and losses (Dice, cross-entropy, class-balanced cross-entropy).
• Trains with Adam optimizer, with tuned epsilon to counteract loss spikes, using 64x64x64 voxel patches.
• Evaluates on the MICCAI 2015 Multi-Atlas Labeling Beyond the Cranial Vault dataset, comparing against external methods.

実験結果

リサーチクエスチョン

• RQ1 Can a TensorFlow-based, modular toolkit provide competitive, reference implementations for common medical image segmentation architectures?
• RQ2 How do data sampling strategies and loss functions affect performance of U-Net and FCN on a standard abdominal CT segmentation task?
• RQ3 What is the relative performance of the DLTK reference implementations compared with published top methods on a public MICCAI challenge dataset?

主な発見

Method	Dataset/Challenge	Dice (DSC)
DLTK U-Net (cross-entropy, class-balanced sampling)	Multi-Atlas Labeling Beyond the Cranial Vault (MICCAI 2015)	81.5
Multi-Atlas method (previous work)	Multi-Atlas Labeling Beyond the Cranial Vault	79.0
Best CNN (previous work)	Multi-Atlas Labeling Beyond the Cranial Vault	75.7

• DLTK U-Net with cross-entropy loss and class-balanced sampling achieved state-of-the-art Dice similarity coefficient on the challenge data (81.5 DSC).
• The U-Net outperformed the FCN across most metrics and organs in their experiments.
• Compared with a multi-atlas segmentation approach, the DLTK U-Net achieved 81.5 DSC versus 79.0 for the multi-atlas method.
• Compared with the best performing CNN reported in prior work, the CNN scored 75.7 DSC on the same data.
• Validation performance reported 78.9 DSC, slightly below the 79.0 value cited for the competing CNN in prior work, suggesting possible overfitting in that prior entry.