Reducing the Model Variance of a Rectal Cancer Segmentation Network

Joohyung Lee, Joohyung Lee 0003, Ji Eun Oh, Min Ju Kim, Bo Yun Hur, Dae Kyung SohnIEEE Access 2019 · DOI

TL;DR

This paper tackles the challenge of overfitting in deep learning models for rectal cancer segmentation, a crucial step in cancer staging. They propose a novel pixel-wise bias-variance analysis to quantify model robustness and introduce two methods – multi-task learning (adding rectum segmentation) and data augmentation (random resizing) – to significantly reduce model variance and improve segmentation accuracy.

What Problem Does This Paper Address?

Accurate segmentation of rectal cancer from MRI scans is vital for precise cancer staging and treatment planning. However, deep learning models, while powerful, are prone to overfitting, especially in medical imaging where annotated data is scarce. Overfitting leads to high model variance, meaning the model's performance fluctuates significantly with different training datasets. This makes the segmentation unreliable and can hinder accurate staging. Existing methods for assessing model robustness often suffer from selection bias or don't focus on specific regions of interest, which is critical for medical image analysis where performance near the tumor is paramount.

Methods

The authors introduce a pixel-wise bias-variance analysis method to quantify and visualize the bias and variance of segmentation networks. This approach allows for the assessment of model robustness within arbitrary regions of interest (ROIs), specifically focusing on the cancerous regions.

To reduce model variance, they propose two key techniques:

  1. Multi-Task Learning (MTL) with Rectum Segmentation: Recognizing the anatomical correlation between rectal cancer and the rectum itself (rectal cancer is typically located within the rectum), they train a network to segment both simultaneously. This "multi-region segmentation network" (MRSN) leverages the shared features and provides rectum boundary information as a byproduct, which is clinically relevant.
  2. Data Augmentation via Random Resizing: To address the inherent scale variations in medical images from different scanners, they employ a data augmentation strategy where mini-batches are resized to random scales during training. This aims to improve the network's scale-robustness.

Datasets

The study utilized a dataset of 1,813 rectal cancer patient MRI scans. From this, 907 images were selected based on specific criteria (e.g., clear T2/T3 stage, no prior treatment). Manual delineations of the rectum and rectal cancer were performed by two gastrointestinal clinical specialists, serving as the ground truth. For bias-variance analysis, 10% of the data was set aside as a test set, and the remaining 90% was used for 9-fold cross-validation to generate nine distinct training sets.

Key Insights and Ideas

  • Geometric Correlation is Key: The anatomical relationship between rectal cancer and the rectum is a strong signal that can be exploited to improve segmentation models.
  • Pixel-wise Analysis for Granular Insights: A pixel-level bias-variance analysis provides a more nuanced understanding of model performance, especially in critical regions, compared to global metrics.
  • MTL for Robustness: Adding a related task (rectum segmentation) to the primary task (rectal cancer segmentation) effectively constrains the model's parameters, leading to better generalization and reduced variance.
  • Scale Augmentation for Real-World Data: Medical images inherently vary in scale due to scanner differences. Random resizing during training is an effective way to make models more robust to these variations.
  • Bias-Variance Trade-off: The study demonstrates that their methods reduce variance without significantly increasing bias, a crucial aspect of improving model performance.

Key Results

  • Reduced Model Variance:
    • Adding rectum segmentation (MRSN) reduced rectal cancer segmentation variance within tumor regions by a factor of 0.90.
  • Data augmentation further reduced variance by a factor of 0.89.
  • These approaches also reduced overall image variance.
  • Improved Segmentation Performance (DSC):
    • MRSN showed statistically significant improvements in tumor DSC and sensitivity compared to the single-region segmentation network (SRSN).
  • MRSN-AUG further improved tumor DSC, rectum DSC, tumor sensitivity, and rectum specificity.
  • Reduced Training Duration:
    • MRSN reduced training duration by a factor of 0.96.
  • Augmentation further reduced MRSN training duration by a factor of 0.78.
  • No Increase in Bias: Neither approach significantly increased model bias.

Limitations and Future Work

  • 2D vs. 3D: The current network operates on 2D images, neglecting information from neighboring slices. The authors suggest developing a 3D segmentation network that utilizes stacked medical images for potentially improved performance.
  • Task-Specific Layer Design: The design of elaborate task-specific layers for both rectum and rectal cancer segmentation was left for future research.
  • DSC as Loss Function: Investigating the use of Dice Similarity Coefficient (DSC) as a loss function for bias-variance decomposition is proposed as an interesting future direction.
  • Generalizability: While the methods are expected to benefit other medical image analysis systems, further validation on different modalities and tasks would be beneficial.

Summary updated

BibTeX
@article{lee2019reducing,
  author = {Joohyung Lee and Joohyung Lee and Ji Eun Oh and Min Ju Kim and Bo Yun Hur and Dae Kyung Sohn},
  title = {{Reducing the Model Variance of a Rectal Cancer Segmentation Network}},
  journal = {IEEE Access},
  year = {2019},
  doi = {10.1109/access.2019.2960371},
}
Figure