Abstract
This study evaluates the effectiveness of traditional handcrafted features—Local Binary Patterns (LBP) and Histogram of Oriented Gradients (HOG)—against deep features extracted from a pre-trained VGG19 network for the classification of histopathology images. Using the KIMIA Path960 dataset, we compare various feature extractors paired with Support Vector Machines (SVM), Decision Trees, and Artificial Neural Networks. Our results demonstrate that LBP features combined with an SVM achieve a superior accuracy of 90.52%, significantly outperforming off-the-shelf deep features (81.14%). These findings highlight the critical role of local texture descriptors in medical imaging, where a significant domain gap exists between natural images and cellular structures.
Introduction
Histopathology, the microscopic examination of tissue to study the manifestations of disease, is fundamental to cancer diagnosis. While deep learning has become the standard in computer vision, its application to medical imaging is often hindered by the "domain gap." Most state-of-the-art models are pre-trained on ImageNet, which contains distinct objects like animals or vehicles. In contrast, histopathology images are characterized by complex, repetitive, and stochastic textures of cells and stroma. This paper investigates whether modern deep features are truly superior to traditional texture-based descriptors for identifying tissue types in medical scans.
Methodology
The research evaluates three distinct feature extraction approaches:
- Local Binary Patterns (LBP): A classic texture descriptor that labels pixels by thresholding the neighborhood and converting the result into a binary number. This method is highly effective at capturing local spatial patterns.
- Histogram of Oriented Gradients (HOG): A descriptor that counts occurrences of gradient orientations in localized portions of an image, typically used for shape and edge detection.
- Deep Features (VGG19): Features extracted from a VGG19 model pre-trained on ImageNet. To identify the optimal level of abstraction, features were extracted from four different depths:
fc1,block5_pool,block4_pool, andblock3_pool.
These features were processed using three classification algorithms: Support Vector Machines (SVM) with an RBF kernel, Decision Trees (DT), and shallow Artificial Neural Networks (ANN). Robustness was ensured through 3-fold cross-validation.
Dataset
The study utilized the KIMIA Path960 dataset, a publicly available benchmark consisting of 960 histopathology images. These images represent 20 different tissue classes, each down-sampled to 308 × 168 pixels. The dataset is particularly challenging due to the high visual similarity between different tissue types, requiring descriptors that can distinguish subtle textural nuances.
Results and Discussion
The experimental results revealed a clear hierarchy in performance:
- Superiority of LBP: The LBP features paired with an SVM achieved the highest overall accuracy of 90.52%. This suggests that local texture information is more discriminative for histopathology than the high-level "object" shapes deep networks are typically optimized to find.
- Deep Feature "Sweet Spot": Among the VGG19 features, the best performance (81.14%) was achieved at the
block4_poollayer. Accuracy dropped significantly at the finalfc1layer (44.79%) and the earlierblock3_poollayer (67.49%). This indicates that as networks go deeper into ImageNet-specific features, they lose the generality required for medical textures. - HOG Failure: HOG performed poorly, peaking at only 36.15% accuracy. Its reliance on edge-orientation is ill-suited for the "blobby" and irregular nature of cellular structures.
Conclusion
Traditional handcrafted features, specifically LBP, remain highly relevant and effective for histopathology image classification. While deep learning offers immense potential, off-the-shelf models pre-trained on natural images may not always outperform specialized texture descriptors. Future work should focus on fine-tuning deep models specifically on medical datasets or developing hybrid models that fuse the local precision of LBP with the hierarchical context of deep neural networks.
Summary updated