[논문 리뷰] Deep learning and its application to medical image segmentation

Holger R. Roth, Chen Shen외 5명|arXiv (Cornell University)|2018. 03. 23.

Radiomics and Machine Learning in Medical Imaging참고 문헌 32인용 수 61

한 줄 요약

이 논문은 복부 CT에서 자동 다기관 분할을 위한 3D 완전 컨볼루션 네트워크(3D U-Net)를 개발하고 위암 CT 데이터셋에서 Dice 점수 거의 최상위 수준을 보고한다.

ABSTRACT

One of the most common tasks in medical imaging is semantic segmentation. Achieving this segmentation automatically has been an active area of research, but the task has been proven very challenging due to the large variation of anatomy across different patients. However, recent advances in deep learning have made it possible to significantly improve the performance of image recognition and semantic segmentation methods in the field of computer vision. Due to the data driven approaches of hierarchical feature learning in deep learning frameworks, these advances can be translated to medical images without much difficulty. Several variations of deep convolutional neural networks have been successfully applied to medical images. Especially fully convolutional architectures have been proven efficient for segmentation of 3D medical images. In this article, we describe how to build a 3D fully convolutional network (FCN) that can process 3D images in order to produce automatic semantic segmentations. The model is trained and evaluated on a clinical computed tomography (CT) dataset and shows state-of-the-art performance in multi-organ segmentation.

연구 동기 및 목표

• Address the challenge of automated semantic segmentation in abdominal CT despite large anatomical variability.
• Propose a 3D fully convolutional network architecture inspired by U-Net for volumetric medical image segmentation.
• Demonstrate end-to-end training with data augmentation and a Dice-based loss to optimize segmentation performance.
• Evaluate the model on a gastric cancer CT dataset across multiple organs and compare to state-of-the-art methods.

제안 방법

• Adopt a 3D FCN with symmetric encoder–decoder (similar to 3D U-Net) using 3x3x3 kernels and 2x2x2 max pooling in the encoder and transposed convolutions in the decoder.
• Incorporate skip connections between encoder and decoder at matching resolutions to preserve high-resolution features.
• Train with randomly cropped subvolumes (64x64x64) on a single GPU, using batch normalization and Adam optimization, with a Dice-based loss for multi-class segmentation.
• Apply data augmentation including smooth B-spline deformations, random rotations and translations to increase robustness and reduce overfitting.
• Process whole volumes at inference time via an overlapping tiles approach and resize outputs to match input dimensions.
• Utilize a voxel-wise softmax to produce per-voxel class probabilities and compute a total loss as a weighted sum of per-class Dice losses (weights set to 1 in this study).

실험 결과

연구 질문

• RQ1 Can a 3D fully convolutional network trained end-to-end achieve accurate multi-organ segmentation in abdominal CT?
• RQ2 What Dice score can be achieved on arteries, veins, liver, spleen, stomach, gallbladder, and pancreas using the proposed 3D FCN?
• RQ3 How do data augmentation and network architecture (including skip connections) affect segmentation performance and robustness?
• RQ4 Is a 3D U-Net-like FCN competitive with other state-of-the-art deep learning approaches for CT organ segmentation across multiple organs?

주요 결과

• Average Dice scores on training data: 89.4% across all organs (per-class variability).
• Average Dice scores on testing data: 89.3% across all organs (per-class variability).
• Organ-wise Dice on testing: artery 83.5%, vein 80.5%, liver 97.1%, spleen 97.7%, stomach 96.1%, gallbladder 85.1%, pancreas 84.9%.
• The model scales to 3D volumetric segmentation with ~19 million parameters and inference under 1 minute per case.
• Data augmentation and volume-based processing help prevent overfitting and enable robust multi-organ segmentation on whole CT volumes.
• The approach yields competitive performance compared to other state-of-the-art architectures for 3D abdominal CT organ segmentation.