Enhancing Histopathological Image Classification via Integrated HOG and Deep Features with Robust Noise Performance

1. Problem Statement

Histopathological image analysis is critical for disease diagnosis but faces significant challenges due to the massive size of Whole Slide Images (WSIs) and the intricate, complex features that are difficult to model with simple mathematical equations. While Deep Learning (DL) has advanced automated diagnostics, existing methods often struggle with:

• Feature Representation: Relying solely on high-level abstract features may miss critical low-level texture and edge information.
• Robustness: Many models degrade significantly when image quality is compromised by noise (low Signal-to-Noise Ratio or SNR), which is common in real-world clinical imaging.
• Performance Gaps: Traditional Machine Learning (ML) techniques often underperform compared to pure DL approaches, yet pure DL models can be brittle under noisy conditions.

The study aims to bridge these gaps by integrating traditional hand-crafted features with deep learning features and evaluating the robustness of these hybrid models against varying noise levels.

2. Methodology

The research utilized the LC25000 dataset, comprising 25,000 color histopathological images across five classes: colon adenocarcinoma, benign colonic tissue, lung adenocarcinoma, lung squamous cell carcinoma, and benign lung tissue.

The experimental framework involved three primary approaches:

A. Fine-Tuned InceptionResNet-v2 as a Classifier

• The pre-trained InceptionResNet-v2 network was fine-tuned for direct classification.
• Architecture: Two custom convolutional blocks were added (128 kernels of 7x7 and 256 filters of 5x5), followed by a fully connected layer.
• Optimization: Stochastic Gradient Descent (SGD) with momentum (0.9) and a piecewise learning rate schedule.

B. Deep Feature Extraction + ML Classifiers

• Feature Extraction: Images were processed through the InceptionResNet-v2 network, and high-level features were extracted from the 'avg pool' layer.
• Augmentation: Random rotations (0°, 90°, 180°, 270°) and translations were applied to increase data diversity.
• Classification: These deep features were fed into five distinct ML models: Decision Tree, Gradient Boosting Machine (GBM), k-Nearest Neighbors (KNN), Neural Network (NN), and Support Vector Machine (SVM).
• Tuning: Hyperparameters were optimized using Bayesian Optimization.

C. Integrated HOG and Deep Features

• Hybrid Approach: Histogram of Oriented Gradients (HOG) features were extracted (cell size 128x128) from images resized to 299x299 pixels.
• Concatenation: HOG features were concatenated with the deep features from InceptionResNet-v2 to create a comprehensive feature vector.
• Classification: The same five ML models were trained on this combined feature set.

D. Robustness Evaluation (SNR Analysis)

• To test resilience, additive Gaussian noise was applied to the dataset at three Signal-to-Noise Ratio (SNR) levels: 40dB (low noise), 35dB (moderate), and 30dB (high noise).
• Model performance was re-evaluated across all three methods under these noisy conditions.

3. Key Contributions

1. Novel Feature Integration: The study introduces a method combining HOG features (capturing edge and texture details) with deep features (capturing high-level abstractions). This hybrid approach leverages the strengths of both traditional computer vision and modern DL.
2. Robustness Benchmarking: Unlike many studies that only report accuracy on clean data, this paper systematically evaluates model resilience against varying SNR levels, identifying which architectures maintain performance in noisy environments.
3. Performance Superiority: The study demonstrates that models trained on deep features (especially when combined with HOG) significantly outperform both the standalone fine-tuned network and models using only traditional features.

4. Key Results

A. Clean Data Performance

• Fine-Tuned Classifier: The standalone fine-tuned InceptionResNet-v2 achieved 96.01% accuracy and 96.8% AUC.
• Deep Features Only: Models trained on deep features significantly outperformed the standalone classifier.
  • Neural Network (NN): Achieved the highest performance with 99.99% AUC and 99.72% accuracy.
  • GBM: 99.77% AUC, 98.05% accuracy.
  • KNN & SVM: Both exceeded 99% AUC.
• Integrated HOG + Deep Features:
  • The combination further enhanced performance for most models.
  • Neural Network: Reached 99.99% AUC and 99.84% accuracy (the peak result of the study).
  • GBM: Improved to 99.9% AUC and 98.73% accuracy.
  • SVM: Achieved 99.96% AUC.

B. Robustness to Noise (SNR Impact)

• Standalone Classifier: Performance collapsed as noise increased, dropping from 80.16% (40dB) to 27.44% (30dB).
• Deep Feature Models: Showed greater resilience.
  • GBM and KNN were the most robust. At 30dB (high noise), GBM maintained >80% accuracy, and KNN maintained ~75%.
  • Neural Network and SVM saw a sharper decline in high-noise environments compared to GBM/KNN.
• Integrated HOG + Deep Features:
  • While performance at 40dB was comparable to deep features alone, the addition of HOG did not significantly improve robustness at lower SNR levels (30-35dB) compared to deep features alone. In some cases, performance was slightly lower than the deep-feature-only approach in high-noise scenarios.

5. Significance and Conclusion

The study concludes that integrating deep learning with traditional feature extraction creates a powerful framework for histopathological classification.

• Optimal Configuration: The Neural Network model trained on combined HOG and deep features yielded the highest accuracy (99.84%) on clean data.
• Noise Resilience: For clinical applications where image quality may vary, GBM and KNN models trained on deep features proved to be the most reliable, maintaining high accuracy even under significant noise (30dB).
• Clinical Impact: The findings suggest that while pure deep learning is powerful, hybrid approaches and specific ML classifiers (like GBM/KNN) offer superior reliability for early disease detection and personalized treatment planning, particularly in non-ideal imaging conditions.

This research provides a robust pathway for developing automated diagnostic tools that are not only highly accurate but also resilient to the practical challenges of real-world digital pathology.