FrGNet: A fourier-guided weakly-supervised framework for nuclear instance segmentation

TL;DR

FrGNet tackles the high annotation burden in nuclear instance segmentation by introducing a Fourier-guided framework that injects a priori nuclear location information through a Fourier Guidance (FG) module and enhances instance representation via a Guide-based Instance Level Contrastive (GILC) module. By fusing Fourier-derived guide masks with a contour-based instance head and applying an instance-level contrastive loss, the approach performs well under both full- and weak-supervised regimes. Empirical results on MoNuSeg and CPM17 show state-of-the-art performance under full supervision and robust performance with limited annotations, with strong generalization to a private dataset without labels. The method reduces labeling requirements while delivering accurate, interpretable segmentation that leverages nuclear characteristics for improved localization and representation. $PQ$, $AJI$, $DQ$, and $SQ$ metrics substantiate the gains, highlighting practical impact for pathology image analysis.

Abstract

Nuclear instance segmentation has played a critical role in pathology image analysis. The main challenges arise from the difficulty in accurately segmenting instances and the high cost of precise mask-level annotations for fully-supervised training.In this work, we propose a fourier guidance framework for solving the weakly-supervised nuclear instance segmentation problem. In this framework, we construct a fourier guidance module to fuse the priori information into the training process of the model, which facilitates the model to capture the relevant features of the nuclear. Meanwhile, in order to further improve the model's ability to represent the features of nuclear, we propose the guide-based instance level contrastive module. This module makes full use of the framework's own properties and guide information to effectively enhance the representation features of nuclear. We show on two public datasets that our model can outperform current SOTA methods under fully-supervised design, and in weakly-supervised experiments, with only a small amount of labeling our model still maintains close to the performance under full supervision.In addition, we also perform generalization experiments on a private dataset, and without any labeling, our model is able to segment nuclear images that have not been seen during training quite effectively. As open science, all codes and pre-trained models are available at https://github.com/LQY404/FrGNet.

Figures (8)

• Figure 1: Motivations. We can get corse nuclear instance segmentation results without any training.
• Figure 2: The workflow of our proposed FrGNet framework. For an input image, we first extract the image features using the feature extraction module to obtain multi-level image features, then use the features maps to generate guide mask and nuclear instances in fourier guidance module. Furthermore, the proposed instance-level comparison module also make use of the feature maps to enhances the feature representation of the nuclear instances.
• Figure 3: Overview of inference stage. For an input image, after processed by FrGNet, the instance head generate original predicted instances, and the fourier guide head generate predicted guide mask of input image. In post process operation, we filter out the instances in original predicted instances group according to the predicted guide mask, producing the final refined instances.
• Figure 4: The flow of guide mask generation. For an image, we use fourier transform, low frequency filtering, inverse fourier transform to get the initial guide mask, which is then normalized and added with binary instance ground truth mask to get the final guide mask which is used to guide the model for training.
• Figure 5: Examples of cropped histopathologic image and corresponding instance mask ground truth.