Partial Labeled Gastric Tumor Segmentation via patch-based Reiterative Learning

Yang Nan, Gianmarc Coppola, Qiaokang Liang, Kunlin Zou, Wei Sun, Dan Zhang, Yaonan Wang, Guanzhen YuarXiv 2017 · arXiv

Abstract

Biomedical image segmentation for gastric cancer often suffers from a lack of complete annotations, as professional labeling is both time-consuming and expensive. This paper presents a reiterative learning framework designed to achieve high-accuracy segmentation using only partially labeled data. By integrating a patch-based Fully Convolutional Network (FCN) with a self-correcting feedback loop and an Overlapped Region Forecast (ORF) algorithm, the proposed method effectively "grows" its own training set. The approach achieved a Mean Intersection over Union (IOU) of 0.883, winning the 2017 China Big Data & Artificial Intelligence Innovation and Entrepreneurship Competition.

Introduction

Identifying the precise extent of gastric tumors in high-resolution scans is a critical yet labor-intensive task. Pathologists often provide "weak" or partial annotations, marking only 20% to 70% of the tumor area. Standard supervised learning models typically treat these missing regions as background (false negatives), which severely degrades performance. The challenge addressed here is training a robust segmentation model when the ground truth is inherently incomplete.

Comparison of original images (a), partial labels (b), and complete ground truth (c).

Methodology

Preprocessing and Architecture

To manage massive 2048x2048 images, the framework utilizes a patch-based approach, slicing images into 512x512 segments. A Flood Fill Algorithm (FFA) is employed to remove background noise and empty tissue spaces. The core architecture is a Fully Convolutional Network (FCN) inspired by U-Net, utilizing Parametric ReLU (PReLU) to maintain gradients and a hybrid loss function combining Binary Cross Entropy and Dice Coefficient loss.

The Fully Convolutional Network (FCN) architecture used for patch-based segmentation.

Overlapped Region Forecast (ORF)

To eliminate boundary errors common in patch-based methods, the authors implemented an ORF algorithm. By using an overlapping sliding window and averaging predictions across multiple views of the same pixel, the model produces smoother, more accurate global segmentations. Reducing the stride (increasing overlap) was found to significantly boost Mean IOU.

Reiterative Learning Loop

The primary innovation is the reiterative learning process. After initial training on partial labels, the model predicts labels for its own training set. Predictions with a confidence threshold greater than 0.7 are merged with the original partial labels to create an "enhanced" dataset. This self-teaching loop allows the model to identify tumor regions missed by human annotators, effectively expanding the training signal.

Results and Discussion

The reiterative framework demonstrated incremental improvements at each stage. The baseline patch model achieved an 85.28% Mean IOU, which rose to approximately 86.4% after the first iteration. The final ensemble model, combining high-recall and high-precision variants, reached a Mean IOU of 0.883 and a Mean Accuracy of 91.09%. This performance secured first place in the 2017 China Big Data competition.

Despite these successes, the model faces limitations in detecting very small cancerous clusters and carries a risk of over-fitting if iterated excessively, as it may begin to reinforce its own errors. Future work will focus on refining confidence thresholds and increasing sensitivity to micro-scale tissues.

Summary updated

BibTeX
@misc{nan2017partial,
  author = {Yang Nan and Gianmarc Coppola and Qiaokang Liang and Kunlin Zou and Wei Sun and Dan Zhang and Yaonan Wang and Guanzhen Yu},
  title = {{Partial Labeled Gastric Tumor Segmentation via patch-based Reiterative Learning}},
  year = {2017},
  eprint = {1712.07488},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1712.07488},
}
Figure