WPMMA-NIS: wavelet pooling and major-minor axis distance monomerization for nuclei instance segmentation in breast whole slide image

This paper proposes WPMMA-NIS (Wavelet Pooling and Major–Minor Axis Distance Monomerization for Nuclei Instance Segmentation in Breast Whole Slide Images), a method that preserves richer detail information for nuclei classification and achieves the monomerization of clustered nuclei. During the feature extraction stage, a wavelet transform-based pooling method is designed to retain more detailed information. In the post-processing stage, a monomerization technique based on the distance between the major and minor axes is devised to separate clustered nuclei. Finally, the monomerized nuclei serve as masks on the pixel-level classification layer, where the predominant pixel class within each nucleus boundary is assigned as its final label. The experimental results demonstrate that WPMMA-NIS achieves performance comparable to state-of-the-art algorithms in pixel-level classification tasks (accuracy: 97.84%, precision: 93.60%, specificity: 99.03%, and recall: 88.74%) and superior performance in object-level multi-class classification tasks (93.11% detection accuracy and 90.18% overall classification accuracy).