Abstract
Medical image analysis has become a topic under the spotlight in recent years. There is a significant progress in medical image research concerning the usage of machine learning. However, there are still numerous questions and problems awaiting answers and solutions, respectively. In the present study, comparison of three classification models is conducted using features extracted using local binary patterns, the histogram of gradients, and a pre-trained deep network. Three common image classification methods, including support vector machines, decision trees, and artificial neural networks are used to classify feature vectors obtained by different feature extractors. We use KIMIA Path960, a publicly available dataset of 960 histopathology images extracted from 20 different tissue scans to test the accuracy of classification and feature extractions models used in the study, specifically for the histopathology images. SVM achieves the highest accuracy of 90.52% using local binary patterns as features which surpasses the accuracy obtained by deep features, namely 81.14%.
Introduction
Histopathology—the microscopic study of diseased tissue—is fundamental to modern oncology and cancer diagnosis. Despite its importance, manual analysis remains a bottleneck: it is labor-intensive, time-consuming, and subject to inter-observer variability. While machine learning has made significant strides in automating medical image analysis, a fundamental question persists: which image representation (feature extraction) method is most effective for the unique visual characteristics of pathology?
Modern computer vision often defaults to "Deep Features" extracted from convolutional neural networks (CNNs) pre-trained on natural image datasets like ImageNet. However, histopathology images differ significantly from natural images; they lack distinct objects like "cars" or "dogs" and instead consist of complex, repetitive cellular textures. This study evaluates whether generic deep features are truly superior to traditional handcrafted descriptors like Local Binary Patterns (LBP) or Histograms of Oriented Gradients (HOG) in this specialized domain.
Methodology
The researchers implemented a comparative framework involving three distinct feature extraction techniques paired with three classification algorithms.
Feature Extraction Techniques
- Local Binary Patterns (LBP): A powerful texture descriptor that characterizes local spatial structures by comparing each pixel with its neighbors to generate a binary code.
- Histogram of Oriented Gradients (HOG): A descriptor that captures edge and shape information by calculating the distribution of intensity gradients.
- Deep Features (VGG19): Features were extracted from a pre-trained VGG19 network. To investigate the impact of layer depth, features were pulled from four different stages: fc1, block5_pool, block4_pool, and block3_pool.
Classification Models
The extracted feature vectors were fed into three standard classifiers: Support Vector Machines (SVM), Decision Trees (DT), and Artificial Neural Networks (ANN). A 3-fold cross-validation approach was employed to ensure the reliability and statistical significance of the results.
Experimental Dataset
The evaluation was performed using the KIMIA Path960 dataset, a benchmark specifically designed for histopathology image retrieval and classification.
- Composition: The dataset contains 960 images.
- Diversity: These images are derived from 20 different tissue scans, representing a wide array of biological patterns including epithelial, connective, and muscle tissues.
- Format: Images are provided in colored TIF format, down-sampled to 308 × 168 pixels.
Results and Discussion
The experimental results revealed a clear performance hierarchy among the feature extraction methods.
Performance Summary
- The Top Performer: The combination of LBP and SVM achieved the highest overall accuracy of 90.52%. This suggests that the texture-centric nature of LBP is exceptionally well-suited for the repetitive patterns found in tissue scans.
- Deep Learning Performance: The best deep feature result was obtained using VGG19 (block4_pool) with an SVM, reaching 81.14%. Notably, features from the middle of the network outperformed those from the final fully connected layers (fc1).
- HOG Limitations: HOG performed poorly, with its best result being only 36.15% (using an ANN). The irregular and dense nature of cellular structures likely resulted in sparse feature vectors that lacked discriminative power.
Key Insights
The study highlights that "texture is king" in pathology. Because histopathology images are defined by cellular distributions rather than macroscopic object shapes, LBP's local comparison logic captures the essential diagnostic information more effectively than generic deep features. Furthermore, the "Deep Dilemma" observed—where middle layers outperformed deeper layers—indicates that the final layers of networks trained on ImageNet become too specialized for natural objects, losing the ability to represent the abstract textures of human tissue.
Conclusion
This research demonstrates that in specialized domains like histopathology, traditional handcrafted features like LBP can still outperform state-of-the-art deep learning features. While deep learning is powerful, its effectiveness is often tied to the relevance of the pre-training data.
Future work should explore fine-tuning deep networks directly on large-scale medical datasets to bridge the performance gap. Additionally, testing these models on "Whole Slide Images" (WSI) would provide further insight into their scalability for clinical environments. For now, LBP remains a robust and computationally efficient choice for pathology image classification.
Summary updated