NoiseAugmentNet-HHO: Enhancing Histopathological Image Classification Through Noise Augmentation
Prem Purusottam Jena, Debahuti Mishra, Kaberi P. Das, Sashikala Mishra · IEEE Access · 2024
Histopathological image analysis is critical in contemporary medical diagnostics, enabling precise examination of tissue samples at microscopic levels. Despite advances in digital pathology and imaging technologies, developing robust deep learning classifiers for histopathological images remains challenging due to inherent variability and complexity. This study investigates the efficacy of noise augmentation and denoising techniques to enhance classifier performance. We systematically evaluate various noise augmentation methods across datasets including BACH, Camelyon17, and BreakHis, aiming to improve robustness under diverse noise conditions such as Salt and Pepper, Speckle, and Poisson. Key innovations include NoiseAugmentNet-HHO, which integrates Harris Hawks optimization (HHO) with VGG16, ResNet50, and deep CNN (DCNN). This model is designed to effectively manage noise types and achieve superior denoising performance, with metrics such as PSNR, entropy, SSIM, MSE, and MAE being used for evaluation. Pixel-wise regression is emphasized for accurate tissue structure assessment and anomaly detection. NoiseAugmentNet-HHO demonstrates significant improvements in classifier performance compared to traditional methods. It achieves high accuracy rates of 94.22% to 98.26% across the BACH, Camelyon17, and BreakHis datasets. Superior sensitivity (up to 98.36%) and specificity (up to 98.10%) further highlight its robust capability in accurately identifying image features. The encoder-decoder architecture of NoiseAugmentNet-HHO ensures robust feature extraction and reconstruction, preserving spatial information essential for accurate classification. Statistical analyses confirm its superiority in maintaining image fidelity and classification accuracy. NoiseAugmentNet-HHO outperforms traditional denoising methods across Salt and Pepper, Speckle, and Poisson noise types, effectively enhancing diagnostic precision. It demonstrates high effectiveness in noise reduction while preserving image details, proving its value in the robust evaluation of histopathological images.