Involution-based efficient autoencoder for denoising histopathological images with enhanced hybrid feature...

Area of Science:

• Medical Imaging
• Computational Pathology
• Artificial Intelligence

Background:

• Histopathology image noise, stemming from hardware, preparation, and environment, hinders disease analysis.
• Existing denoising models often fail to capture spatial features and are computationally intensive.
• Sparse spatial features in histopathology images are critical for diagnosis but can be degraded by denoising.

Purpose of the Study:

• To develop an efficient and precise denoising method for histopathology images.
• To address the limitations of current models in extracting spatial features and computational cost.
• To create a lightweight autoencoder capable of preserving critical diagnostic information.

Main Methods:

• Proposed a lightweight autoencoder (43.11 KB) by integrating an involution layer into a small convolutional model.
• Employed a hybrid approach for channel-specific and location-specific feature extraction.
• Evaluated performance using SSIM Loss and Peak-Signal-to-Noise-Ratio on Malaria Blood Smear and CRC datasets.

Main Results:

• Achieved low SSIM Loss (0.058 and 0.34) when denoising images with Gaussian noise (0.3).
• The autoencoder has minimal weight parameters (11,037) and FLOPs (81,630,000).
• Demonstrated over 20x greater computational efficiency compared to the next best model (Xception).

Conclusions:

• The proposed lightweight autoencoder offers superior denoising performance for histopathology images.
• The model provides significant reductions in memory and computational overhead.
• This method represents a highly efficient solution for histopathology image denoising, minimizing artifact generation.