FEEDNet: a feature enhanced encoder-decoder LSTM network for nuclei instance segmentation for histopathological...

Area of Science:

• Computational pathology
• Medical image analysis
• Deep learning for cancer diagnosis

Background:

• Accurate cell nuclei segmentation is crucial for cancer diagnosis using whole slide images (WSIs).
• Challenges in segmenting nuclei from hematoxylin and eosin (HE) stained WSIs include noise and uneven staining.
• Existing methods often struggle with preserving location information and handling multi-channel data effectively.

Purpose of the Study:

• To propose a novel deep learning model, FEEDNet, for accurate nuclei segmentation in HE-stained WSIs.
• To address the limitations of current segmentation techniques by incorporating feature enhancement and LSTM units.
• To develop a model suitable for both research and deployment on memory-constrained devices.

Main Methods:

• Developed FEEDNet, an encoder-decoder network featuring "feature enhancement blocks" (FE-blocks) and LSTM units.
• FE-blocks preserve pixel intensities by concatenating downsampled images, avoiding pooling layer information loss.
• Employed a multiclass segmentation approach for datasets with class information to generate improved binary masks.

Main Results:

• FEEDNet achieved state-of-the-art panoptic quality (PQ) on the CoNSeP and CPM-17 datasets, and the second-best PQ on the Kumar dataset.
• The model size was significantly reduced from 64.90 MB (32-bit float) to 16.51 MB (INT8 quantization) with minimal performance degradation.
• Demonstrated generalized class-aware binary segmentation accuracy across multiple datasets.

Conclusions:

• FEEDNet offers a highly accurate and efficient solution for nuclei segmentation in histopathology.
• The model's reduced size makes it practical for deployment on edge devices with limited memory.
• The proposed method advances automated cancer diagnosis through improved image analysis.