Automatic classification of cells in microscopic fecal images using convolutional neural networks

Xiaohui Du, Lin Liu, Xiangzhou Wang, Guangming Ni, Jing Zhang, Ruqian Hao, Juanxiu Liu, Yong LiuBioscience 2019 · DOI

TL;DR

This paper presents a novel deep learning approach for automatically identifying and classifying red blood cells (RBCs), white blood cells (WBCs), and molds in microscopic fecal images. By combining a morphology-based region proposal method with a PCA-enhanced Inception-v3 convolutional neural network, the system achieves high accuracy and efficiency, paving the way for automated clinical diagnostic tools.

What Problem Does This Paper Address?

Microscopic examination of fecal samples is crucial for clinical diagnosis, involving the manual counting and identification of various cell types like RBCs, WBCs, and molds. This manual process is time-consuming, prone to human error, and lacks speed and precision, especially in smaller healthcare facilities. While automated vision-based methods exist, they often focus on single targets, leading to low overall detection efficiency and accuracy. The challenge lies in developing a robust system that can accurately identify and classify multiple cell types within complex fecal backgrounds, which are often cluttered with impurities that can resemble target cells.

Methods

The proposed algorithm consists of two main stages:

  1. Region Proposal (Candidate Segmentation):
    • A morphology-based approach is used to extract candidate regions from complex backgrounds. This involves pre-processing the image using Sobel operators in four orientations to create a marginalized image, followed by a local mean thresholding to generate a binary image.
  2. Connecting component analysis is then applied, with regions filtered based on area and bounding box dimensions to remove noise and irrelevant objects.
  3. This method is compared against Selective Search (SS), a popular region proposal algorithm, demonstrating superior performance in terms of lower missing detection rates and significantly faster processing times.
  4. Candidate Recognition (Object Recognition):
    • A novel convolutional neural network (CNN) architecture, PCA-Inception-v3, is employed for classifying the identified candidates.
  5. The Inception-v3 network is used for feature extraction from the cropped candidate regions.
  6. Principal Component Analysis (PCA) is then applied to reduce the dimensionality of the extracted feature vectors, which helps in unifying feature dimensions and retaining crucial information.
  7. The dimensionality-reduced features are fed into a classification network (for type recognition) and a regression network (for location correction).
  8. The model is trained in two modules: Module 1 uses transfer learning with Inception-v3 pre-trained on ImageNet, and Module 2 uses the PCA-compressed features for final classification and regression.

Datasets

  • Training and Evaluation Data: The study utilized 17,933 stool samples collected from the Sixth People's Hospital of Chengdu, Sichuan Province. These samples were processed to yield a total of 89,665 microscopic images.
  • Ground Truth: After manual annotation by clinical doctors, the dataset contained ground-truth labels for 4,459 RBCs, 4,305 WBCs, and 6,536 molds.
  • Dataset Split: A 4:1:1 split was used for training, validation, and testing sets, with cross-validation employed for performance evaluation.

Key Insights and Ideas

  • Morphology for Complex Backgrounds: The paper highlights that traditional morphological methods, while useful, struggle with complex backgrounds and impurities. Their proposed morphology-based region proposal, enhanced with specific operators and thresholding, effectively handles these challenges.
  • PCA for Feature Dimensionality Reduction: Integrating PCA with Inception-v3 is a key innovation. It addresses the issue of varying input sizes for CNNs and reduces feature dimensionality without significant loss of information, improving efficiency and potentially generalization.
  • Two-Module Training Strategy: The split training approach (transfer learning for Module 1, PCA-compressed features for Module 2) is a practical solution to overcome the limitations of end-to-end training when PCA is involved.
  • Efficiency Gains: The combination of a faster region proposal method and an optimized CNN architecture leads to a significant reduction in processing time compared to other methods.

Key Results

  • High Average Precision: The proposed PCA-Inception-v3 model achieved an average Precision of 90.7% across all cell types.
  • Efficient Detection: The total time for detecting and classifying cells in an image (1600x1200 resolution) was approximately 1200 ms.
  • Improved Region Proposal: The morphology-based region proposal method demonstrated a lower missing detection rate (1.3%) and was six times faster than Selective Search (SS).
  • Comparative Performance: The proposed model outperformed other mainstream CNN models like VGG-19, Inception-v3, Inception-v4, and Inception-Resnet-v2 in terms of Precision and Recall when tested on the extracted candidate sets.

Limitations and Future Work

  • Overfitting Concerns with Small Datasets: The paper acknowledges that deep learning methods, including those used here, can be prone to overfitting when sample sizes are small. While data augmentation is mentioned as a solution, the challenge of small sample sizes remains a general concern in deep learning.
  • Subjectivity in Manual Annotation: While the study relies on expert annotation for ground truth, any manual process inherently carries a degree of subjectivity.
  • Integration into Clinical Devices: The authors suggest that the algorithm can be integrated into automatic fecal detection systems and micro-imaging intelligent devices, implying future work in hardware integration and real-world clinical deployment.
  • Further Model Optimization: While the PCA-Inception-v3 model shows excellent results, further exploration of CNN architectures and optimization techniques could potentially yield even better performance.

Summary updated

Abstract

The analysis of fecal-type components for clinical diagnosis is important. The main examination involves the counting of red blood cells (RBCs), white blood cells (WBCs), and molds under the microscopic. With the development of machine vision, some vision-based detection schemes have been proposed. However, these methods have a single target for detection, with low detection efficiency and low accuracy. We proposed an algorithm to identify the visible image of fecal composition based on intelligent deep learning. The algorithm mainly includes region proposal and candidate recognition. In the process of segmentation, we proposed a morphology extraction algorithm in a complex background. As for the candidate recognition, we proposed a new convolutional neural network (CNN) architecture based on Inception-v3 and principal component analysis (PCA). This method achieves high-average <i>Precision</i> of 90.7%, which is better than the other mainstream CNN models. Finally, the images within the rectangle marks were obtained. The total time for detection of an image was roughly 1200 ms. The algorithm proposed in the present paper can be integrated into an automatic fecal detection system.
BibTeX
@inproceedings{du2019automatic,
  author = {Xiaohui Du and Lin Liu and Xiangzhou Wang and Guangming Ni and Jing Zhang and Ruqian Hao and Juanxiu Liu and Yong Liu},
  title = {{Automatic classification of cells in microscopic fecal images using convolutional neural networks}},
  booktitle = {Bioscience},
  year = {2019},
  doi = {10.1042/bsr20182100},
}
Figure