Classification of Hematoxylin and Eosin Images Using Local Binary Patterns and 1-D SIFT Algorithm

Oguzhan Oguz, A. Enis Çetin, Rengül Çetin-AtalayIWCIM@EUSIPCO 2017 · OpenReview

TL;DR

This paper tackles the challenge of automatically classifying H&E stained liver tissue images as cancerous or normal. By combining Local Binary Patterns (LBP) with a novel feature extraction method for the 1-D SIFT algorithm, they achieve an impressive 88% classification accuracy, offering a promising automated tool for pathologists.

What Problem Does This Paper Address?

Cancer is a major global health concern, and early detection is crucial. Histopathological analysis of tissue samples, particularly using Hematoxylin and Eosin (H&E) staining, is a cornerstone of cancer diagnosis. However, this process is time-consuming and relies heavily on the expertise of pathologists. The paper aims to automate a part of this process by developing a computational method to distinguish between normal and cancerous liver tissue in H&E stained images. This could potentially speed up diagnosis and reduce the burden on human experts.

Methods

The core of this work lies in the fusion of two feature extraction techniques: Local Binary Patterns (LBP) and a modified One-Dimensional Scale Invariant Feature Transform (1-D SIFT).

  • Local Binary Patterns (LBP): This is a well-established texture descriptor. It works by comparing the grayscale value of a central pixel with its neighbors. A binary code is generated based on whether neighbors are brighter or darker than the center. These codes are then aggregated into a histogram, providing a representation of local image patterns. The paper explores different approaches to extract LBP histograms, including using a grid-based method and extracting from the entire image. They also incorporate uniform LBP patterns, which are known to be more robust.
  • 1-D SIFT Algorithm: The paper adapts the Scale Invariant Feature Transform (SIFT) for 1-D data, specifically applied to color histograms. The novelty here is a new feature vector extraction process. Instead of directly using gradient magnitudes as in the original SIFT, they propose two methods:
  • Standard 1-D SIFT: This involves extracting key points from color histograms and then creating feature vectors based on the gradient values at these key points.
  • Magnitude 1-D SIFT (M-1-D SIFT): This is the proposed enhancement. It utilizes the magnitudes of the LBP histogram itself as features, rather than gradients. This approach aims to capture more direct information about the texture patterns.

The LBP histograms are concatenated with color histograms to provide a richer feature set before being fed into the 1-D SIFT algorithms for classification.

Datasets

The study utilized a dataset of 454 H&E stained liver images. These images were sourced from 56 different patients, with 270 images classified as cancerous and 184 as normal. The images were acquired at 20x magnification and resized to 300x300 pixels. The dataset was obtained from Biomax.

Key Insights and Ideas

  • Fusion Power: Combining LBP and 1-D SIFT, especially with the proposed M-1-D SIFT, proves to be more effective than using either method alone. This highlights the benefit of integrating texture and scale-invariant feature extraction.
  • LBP Histogram Granularity Matters: Extracting LBP histograms using a grid-based approach (Experiment I & II) or from the whole image (Experiment IV) yields different results. The grid approach, when combined with color histograms (Experiment III), appears to be particularly beneficial.
  • Magnitude is Key for SIFT Adaptation: The novel M-1-D SIFT approach, which uses histogram magnitudes directly, significantly outperforms the standard 1-D SIFT that relies on gradients. This suggests that for this specific application, the raw intensity distribution within the histogram is more discriminative.
  • Color Adds Crucial Information: Incorporating color histograms alongside LBP histograms substantially boosted classification accuracy. This indicates that color variations, beyond just texture, play a vital role in distinguishing cancerous from normal liver tissue in H&E images.

Key Results

The paper reports classification accuracies achieved using two methods: Keypoint Matching (KM) and Efficient Nearest Neighbor Indexing (ENNI). The best performance was achieved with Experiment IV, which combined the M-1-D SIFT algorithm with uniform LBP histograms extracted from the whole image and color histograms. This experiment yielded an accuracy of 88.12% (KM) and 88.10% (ENNI). Experiment III, which used a grid-based LBP histogram extraction, also performed well, achieving around 82.15% accuracy. The M-1-D SIFT approach consistently outperformed the standard 1-D SIFT.

Limitations and Future Work

  • Limitations: The paper acknowledges that while there are obvious pattern differences between normal and cancerous images, some similarities also exist, making classification challenging. The dataset size, while decent, could potentially be expanded. The paper doesn't explicitly detail the computational cost of the proposed methods.
  • Future Work: The authors plan to explore different types of LBP histograms and conduct additional experiments. They also aim to further extend the M-1-D SIFT feature extraction process, suggesting there's room for further optimization and refinement of their approach.

Summary updated

Abstract

In this paper, Hematoxylin and Eosin (H&E) stained liver images are classified by using both Local Binary Patterns (LBP) and one dimensional SIFT (1-D SIFT) algorithm. In order to obtain more meaningful features from the LBP histogram, a new feature vector extraction process is implemented for 1-D SIFT algorithm. LBP histograms are extracted with different approaches and concatenated with color histograms of the images. It is experimentally shown that, with the proposed approach, it possible to classify the H&E stained liver images with the accuracy of 88 % .
BibTeX
@inproceedings{oguz2017classification,
  author = {Oguzhan Oguz and A. Enis Çetin and Rengül Çetin-Atalay},
  title = {{Classification of Hematoxylin and Eosin Images Using Local Binary Patterns and 1-D SIFT Algorithm}},
  booktitle = {IWCIM@EUSIPCO},
  year = {2017},
  url = {https://openreview.net/forum?id=00eVvqWkbk},
}
Figure