Enforcing 3D Coherence in Semi-Supervised Segmentation for Pancreatic Tumor Histopathology from Light Sheet Fluorescence Microscopy

Yousif Hashisho, Diana Pinkert-Leetsch, Jeannine Missbach-Guentner

Keywords: Semi-Supervised Learning, Light Sheet Fluorescence Microscopy, Sparse Annotation, 2D and 3D Segmentation, Pancreatic Cancer, Continuity Loss

Abstract: Light sheet fluorescence microscopy (LSFM) provides unprecedented two-dimensional (2D) tomographic views and three-dimensional (3D) reconstructions of tissue volumes, but generates such large data sets that complete annotation is not feasible. This results in volumes with sparse axial annotations, where ground truth is available for only a small fraction of slices. Standard semi-supervised learning (SSL) methods often fail in this regime, unable to bridge the large gaps between labeled slices to produce coherent 3D segmentations. To address this, we propose a novel SSL framework designed to enforce 3D anatomical plausibility from sparse 2D supervision. The core of our contribution is an axial continuity loss, a regularization term that enforces prediction consistency between adjacent unlabeled slices. This loss is integrated into a voxel-aware Mean-Teacher framework that effectively leverages abundant unlabeled data. We validate our approach on a 3D LSFM dataset of human pancreatic ductal adenocarcinoma (PDAC), which we collected and sparsely annotated for this study. Our experiments show that standard SSL baselines degrade in performance as annotations become sparser, producing noisy predictions between labeled slices. In contrast, our full framework, which integrates an attention-gated 3D U-Net with our proposed continuity loss, maintains robust 3D coherence even in low-data regimes, enabling reliable histopathological analysis from minimal annotations.

Primary Subject Area: Learning with Noisy Labels and Limited Data

Secondary Subject Area: Segmentation

Registration Requirement: Yes

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 277