Texture Feature-Based Classification on Transrectal Ultrasound Image for Prostatic Cancer Detection

Xiaofu Huang, Ming Chen, Peizhong Liu, Yongzhao Du, Lei ChenComput. Math. Methods Medicine 2020 · OpenReview

TL;DR

This paper introduces a novel texture feature fusion method for classifying prostate cancer from transrectal ultrasound (TRUS) images. By combining local binary patterns (LBP) and Gaussian Markov random fields (GMRF) with optical density preprocessing, the method achieves promising accuracy in distinguishing cancerous from non-cancerous tissues, offering a potential improvement over subjective visual interpretation.

What Problem Does This Paper Address?

Prostate cancer is a significant health concern, and early detection is crucial for effective treatment. Transrectal ultrasound (TRUS) is a common imaging modality for guiding prostate biopsies, but its visual interpretation by clinicians is often subjective and can lead to missed diagnoses or overdiagnosis. While TRUS is cost-effective and non-invasive, the inherent speckle noise, low contrast, and artifacts in ultrasound images make it challenging to accurately identify cancerous lesions based on texture alone. Existing methods often struggle with the subtle differences in texture between healthy and cancerous prostate tissue, necessitating improvements in automated analysis to aid clinicians and reduce the burden of invasive biopsies.

Methods

The proposed method tackles the challenge of TRUS image analysis through a multi-stage approach:

  • Image Preprocessing: To enhance image quality and mitigate noise, the authors employ an optical density conversion technique. This process converts the grayscale intensity of pixels into optical density values, which are then linearly mapped to an 8-bit depth image. This step aims to improve contrast and clarify image details, making subsequent texture analysis more effective.
  • Texture Feature Extraction: Two complementary texture feature extraction methods are utilized:
  • Local Binary Patterns (LBP): This method describes local texture features by comparing the gray value of a central pixel with its eight neighbors. It generates a binary code that captures local texture information. The authors use an improved LBP operator to address rotation invariance.
  • Gaussian Markov Random Field (GMRF): This probabilistic model describes image structure and texture by considering the correlation between a pixel and its surrounding pixels. It models the local conditional probability density of pixel values, providing a way to capture more complex textural patterns.
  • Feature Fusion: The extracted LBP and GMRF features are then fused. The paper highlights that combining different feature extraction methods can often lead to higher classification accuracy than using individual features alone. The fusion strategy involves a linear combination of the two feature sets.
  • Classification: The fused texture features are fed into a Support Vector Machine (SVM) classifier. SVM is chosen for its effectiveness in pattern recognition, particularly with high-dimensional and potentially small datasets. The SVM is trained to distinguish between positive (cancerous) and negative (normal) samples.

Datasets

The study utilized a dataset of 342 transrectal ultrasound images obtained from Zhangzhou Hospital affiliated to Fujian Medical University. These images were collected between March 2019 and November 2019. The data was categorized based on pathological biopsy results.

  • Training Set: 36 cases (256 images) - 18 positive (130 images) and 18 negative (126 images).
  • Test Set: 12 cases (86 images) - 6 positive (40 images) and 6 negative (46 images).

The authors note that the data is available under license from the hospital and cannot be made freely available.

Key Insights and Ideas

  • Complementarity of Texture Features: The core insight is that combining different texture analysis techniques (LBP and GMRF) provides a more comprehensive representation of prostate tissue than either method alone. This fusion captures both local patterns and spatial correlations, leading to improved classification.
  • Importance of Preprocessing: Optical density conversion significantly enhances image contrast and detail, which is crucial for accurate texture feature extraction from inherently noisy ultrasound images.
  • SVM's Suitability: SVM proves to be an effective classifier for this task, demonstrating its robustness in handling the complex feature space generated by texture analysis.
  • Addressing Subjectivity: The proposed automated method offers a potential solution to the subjectivity inherent in manual interpretation of TRUS images, aiming for more consistent and objective diagnoses.

Key Results

The proposed texture feature fusion method achieved the following performance metrics on the test set:

  • Accuracy (ACC): 70.93%
  • Sensitivity (SEN): 70.00%
  • Specificity (SPEC): 71.74%

These results are reported to be superior to those obtained using individual texture features (LBP, GMRF, GLCM, HOG, GLDS) and also outperform other common classifiers like KNN, Decision Trees, and Random Forests when applied to the fused features. The paper also highlights that preprocessing improved the classification accuracy of individual features.

Limitations and Future Work

The authors acknowledge several limitations and suggest avenues for future research:

  • Incomplete Labelling: A significant challenge is the difficulty in precisely mapping histological findings to pixel-level labels in ultrasound images. The study uses an "imperfect label" where positive samples might contain normal tissue, though negative samples are reliably normal. This imperfect annotation can affect classification accuracy.
  • Data Scarcity: While 342 images were used, the authors recognize that obtaining more data, especially from multiple hospitals, is crucial for further improving the method's robustness and generalizability.
  • Real-time Application: Developing a real-time image recognition system for TRUS is identified as a future goal, which will require further optimization and potentially more advanced deep learning approaches.
  • Disease Heterogeneity: The authors suggest that prostate cancer and normal prostate tissue might have different underlying structures, which could be further explored to improve classification.
  • Potential for Deep Learning: While not explicitly detailed in this paper, the challenges of real-time recognition and the need for more sophisticated feature extraction hint at the potential for deep learning architectures in future work.

Summary updated

BibTeX
@inproceedings{huang2020texture,
  author = {Xiaofu Huang and Ming Chen and Peizhong Liu and Yongzhao Du and Lei Chen},
  title = {{Texture Feature-Based Classification on Transrectal Ultrasound Image for Prostatic Cancer Detection}},
  booktitle = {Comput. Math. Methods Medicine},
  year = {2020},
  url = {https://openreview.net/forum?id=aZYoPCTBcl},
}
Figure