PaperlayerView exact revision on arXiv

Source-linked AI summary

An attempt at beating the 3D U-Net

Fabian Isensee, Klaus H. Maier-Hein

arXiv:1908.02182v2eess.IVcs.CVcs.LG

TL;DR

Before KiTS, no public dataset with kidney-tumor labels was available, and relatively few segmentation algorithms targeted kidney tumors. The authors compare U-Net-inspired models, some with encoder residual blocks, in KiTS 2019 and submit their best cross-validation U-Net model. 91.23 Composite Dice; the method beat all 105 competing teams and won KiTS 2019 by 0.6 Composite Dice over second place.

  • Problem

    Before KiTS, no public dataset with kidney-tumor labels was available, and relatively few segmentation algorithms targeted kidney tumors.

    Source: §1 · p. 1

  • Method

    The authors compare U-Net-inspired models, some with encoder residual blocks, in KiTS 2019 and submit their best cross-validation U-Net model.

    Source: §1 · p. 2

  • Results

    91.23 Composite Dice; the method beat all 105 competing teams and won KiTS 2019 by 0.6 Composite Dice over second place.

    Source: §4 · p. 6

  • Takeaways & Limitations

    The residual 3D U-Net was chosen for test prediction because its scores were marginally higher, despite no clear winner in the architecture comparison.

    Source: §4 · p. 6

    The architecture comparison used a practical 12 GB GPU memory constraint rather than holding parameter or layer counts constant.

    Source: §4 · p. 6

Abstract

from arXiv · show

The U-Net is arguably the most successful segmentation architecture in the medical domain. Here we apply a 3D U-Net to the 2019 Kidney and Kidney Tumor Segmentation Challenge and attempt to improve upon it by augmenting it with residual and pre-activation residual blocks. Cross-validation results on the training cases suggest only very minor, barely measurable improvements. Due to marginally higher dice scores, the residual 3D U-Net is chosen for test set prediction. With a Composite Dice score of 91.23 on the test set, our method outperformed all 105 competing teams and won the KiTS2019 challenge by a small margin.

1 Introduction

Kidney-tumor segmentation is proposed as a step toward better treatment planning and research, but public kidney-tumor labels were unavailable before KiTS. The paper tests U-Net architectural variants in KiTS 2019 against a backdrop of uncertainty about whether architectural changes improve medical segmentation.

1 Introduction

  • Kidney and tumor segmentation could support surgery planning and research relating tumor morphology to surgical outcomes.Source: §1 · p. 1
  • KiTS addressed the lack of public kidney-tumor labels by providing 210 annotated CT scans for training and 90 for testing.The paper notes that few algorithms had been designed specifically for kidney-tumor segmentation.Source: §1 · p. 1
  • Prior work reported strong plain U-Net results [7], raising questions about the need for extensive architectural modifications.Medical 3D segmentation commonly uses U-Net variants with residual or dense connections.Source: §1 · p. 1
  • The study tests different U-Net-inspired models, some with residual blocks in the encoder, in the KiTS 2019 challenge; its submission uses the best U-Net model based on training-set cross-validation.Source: §1 · p. 2

2 Method

The method compares plain, residual, and pre-activation residual 3D U-Nets for KiTS kidney and tumor segmentation. It resamples and normalizes CT scans, trains the models with patch-based supervision, and modifies selected training labels after manual review.

2.2 Network architecture

  • The experiments compare plain, residual, and pre-activation residual 3D U-Nets, with residual blocks in the encoder for the latter two variants.The residual model uses fewer initial feature maps to accommodate its larger memory footprint; the pre-activation variant shifts normalization and ReLU before convolutions.Source: §2 · p. 2, §2.2 · pp. 3–4, Figure 1 · p. 3

2 Method

  • The paper develops and trains three U-Net-inspired architectures for the KiTS challenge.Source: §2 · p. 2

2.1 Preprocessing

  • All cases were resampled to 3.22 × 1.62 × 1.62 mm spacing, with a median training-image shape of 128 × 248 × 248 voxels, and intensities were clipped and normalized.The intensity range was clipped to [−79, 304], then shifted by 101 and divided by 76.9.Source: §2.1 · p. 2

2.3 Network training

  • Models were trained for 1,000 epochs using batch size 2, 80 × 160 × 160 patches, and a summed cross-entropy and Dice loss.Training included multiresolution supervision and scaling, rotations, brightness, contrast, gamma, and Gaussian-noise augmentation.Source: §2.3 · p. 4

2.4 Dataset modifications

  • After reviewing cases with the worst tumor Dice, the authors excluded four disputed cases and replaced two confirmed faulty annotations with network-generated segmentations.The replacements were visually inspected without manual fine-tuning, and the challenge organizers permitted excluding or modifying training cases.Source: §2.4 · p. 4

3 Results

The three models produced remarkably similar cross-validation results, leaving the authors without a confident winner; they selected the residual 3D U-Net for test prediction. The submitted method then won the KiTS 2019 challenge.

3 Results

  • Cross-validation results for the three models were so similar that the authors did not feel comfortable declaring one better.These five-fold scores used a slightly modified training dataset and may not be comparable with other challenge submissions.Source: §3 · p. 5, Table 1 · p. 5
  • Ensembling did not improve segmentation quality, while the residual 3D U-Net had the highest mean and tumor Dice scores and was selected for test prediction.The challenge score is based on the mean Dice score between kidney and tumor.Source: §3 · p. 5
  • The method won KiTS 2019, outperforming all competing methods on the test set.Table 2 presents the top five teams and ranks them by Composite Dice.Source: §3 · p. 5, Table 2 · p. 6

4 Discussion

The residual 3D U-Net won KiTS2019 by a small margin, although experiments did not establish a clear architectural winner. The discussion emphasizes hardware-constrained comparison and the need for stronger evaluation.

  • 91.23 Composite Dice on the test set beat all 105 competing teams, winning KiTS2019 by 0.6 Composite Dice over second place.Source: §4 · p. 6
  • Cross-validation comparisons were inconclusive, and ensembling the three models did not improve on the best single-model result.Source: §4 · p. 6
  • Poor residual-network implementation or hyperparameter choices could explain the lack of improvement, and the comparison was not tested for statistical significance.Source: §4 · p. 6
  • An implementation error gave the plain U-Net leaky ReLUs and residual models ReLUs; the authors did not expect this to affect results or conclusions.Source: §4 · p. 6
  • Architectures were compared under a 12 GB Titan X GPU memory constraint rather than equal parameter or layer counts.Source: §4 · p. 6
  • Across tested, reasonable-biased target-spacing and patch-size configurations, results were basically identical, suggesting robustness to these choices on this dataset.Source: §4 · p. 7
  • Future work should use extensive architecture-specific hyperparameter optimization and thorough statistical analysis to evaluate differences between plain and modified U-Nets.Source: §4 · p. 7

Related papers

Loading 1908.02182v2…