FEEDNet

FEEDNet segments cell nuclei in hematoxylin and eosin (HE)-stained whole slide images (WSIs) to improve segmentation accuracy in the presence of noise-induced intensity variations and uneven staining.

Key Features:

• Encoder–Decoder with LSTM units: FEEDNet employs an encoder–decoder architecture augmented with Long Short-Term Memory (LSTM) units to capture compact multi-channel representations for histopathological images.
• Feature Enhancement Blocks (FE-blocks): FE-blocks preserve spatial location information lost during pooling by concatenating a downsampled version of the original image to maintain pixel intensities and enhance feature representation.
• Class-aware binary segmentation: A generalized class-aware approach trains a multiclass segmentation model that outputs per-class masks and derives refined binary nuclei masks leveraging available class labels.
• Evaluation and model compactness: Evaluated on CoNSeP, Kumar, and CPM-17 datasets, achieving best Panoptic Quality (PQ) on CoNSeP and CPM-17 and second-best PQ on Kumar; model size is 64.90 MB (FP32) and 16.51 MB with INT8 quantization, with quantization reported to not substantially compromise predictive performance.

Scientific Applications:

• Histopathological nuclei segmentation: Produces instance-level nuclei masks from HE-stained WSIs to support quantitative tissue analysis in cancer diagnosis and other disease studies.

Methodology:

FEEDNet integrates Long Short-Term Memory (LSTM) units within an encoder–decoder network, uses Feature Enhancement Blocks that concatenate a downsampled original image to retain spatial and intensity information, and implements a generalized class-aware multiclass segmentation output to derive binary nuclei masks.