Introducing Hann windows for reducing edge-effects in patch-based image segmentation.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Computationally intensive methods like Convolutional Neural Networks (CNNs) face limitations with large image sizes.
• Biological and medical imaging often produce gigapixel images requiring segmentation into smaller patches for processing.
• Processing image patches can introduce undesirable artifacts, such as edge effects, in the final reconstructed image.

Purpose of the Study:

• To introduce and evaluate windowing methods from signal processing to mitigate edge artifacts in CNN-based image segmentation.
• To compare the effectiveness of different weighting strategies for overlapping image patches.

Main Methods:

• Implemented overlapping patch reconstruction with 2D window weighting.
• Compared simple averaging with Hann, Bartlett-Hann, Triangular, and Cui et al. windows.
• Evaluated performance using Structural Similarity Index (SSIM) and Dice score.

Main Results:

• The cosine-based Hann window demonstrated the best performance improvement, as measured by SSIM.
• The Dice score indicated a reduction in classification errors near patch edges.
• The proposed windowing method is compatible with any CNN segmentation model without modification.

Conclusions:

• Windowing methods effectively reduce edge artifacts in CNN-based image segmentation of large images.
• The Hann window offers a superior approach for reconstructing predictions from overlapping image patches.
• This technique significantly enhances the accuracy of CNN predictions in image segmentation tasks.