Scribble-based fast weak-supervision and interactive corrections for segmenting whole slide images

TL;DR

This work tackles the problem of segmenting whole-slide histopathology images with limited pixel-accurate annotations by introducing a fast, scribble-based weakly supervised framework that enables a pathologist to iteratively refine results with minimal input. A patch-level classifier is trained on scribble-derived labels to produce an initial rough segmentation, which is then refined via two fast correction strategies: a naïve latent-space SVM approach and an uncertainty-based variant that uses Monte Carlo dropout to adapt correction strength and guide scribble placement with a signed uncertainty map. On Camelyon16 data, the uncertainty-based method achieves all main metrics above $90 ext{ extpercent}$ after a few correction passes and outperforms the naïve approach by roughly $17 ext{ extpercent}$ in efficiency, demonstrating the practical viability of a real-time, human-in-the-loop segmentation workflow. The approach provides interpretable uncertainty guidance and maintains rapid update times ($ ext{≤}1 ext{s}$ per correction), supporting clinical use with minimal annotation burden.

Abstract

This paper proposes a dynamic interactive and weakly supervised segmentation method with minimal user interactions to address two major challenges in the segmentation of whole slide histopathology images. First, the lack of hand-annotated datasets to train algorithms. Second, the lack of interactive paradigms to enable a dialogue between the pathologist and the machine, which can be a major obstacle for use in clinical routine. We therefore propose a fast and user oriented method to bridge this gap by giving the pathologist control over the final result while limiting the number of interactions needed to achieve a good result (over 90\% on all our metrics with only 4 correction scribbles).

Figures (5)

• Figure 1: Patch extraction along scribbles created for the 2 classes: tumoral and non-tumoral tissues.
• Figure 2: Scribble creation process.
• Figure 3: Incremental segmentation correction process.
• Figure 4: WSI-level uncertainty measures versus $F_1$ scores of the VGG16 compared with the ground truth calculated on all the WSI patches.
• Figure 5: Visualization of the signed uncertainty map with $t_{thresh} = 0.33$ (left) and heatmap (right) of one WSI. Zoomed regions show areas with TPs (bottom right)and FPs (top right). The signed uncertainty map shows high uncertainty values in the FP region and rather low values in the TP one.