Background In routine histopathology, macroscopic surgical specimens are dissected into multiple microscopic tissue fragments that are processed and analyzed independently. This fragmentation disrupts spatial continuity, making it difficult to integrate local findings into a coherent global view of tumor extent, orientation, and resection margins. Especially for malignant disease or malignoma, this is an issue of particular relevance forwhere positive margins are associated with increased mortality. Reconstructing the original macroscopic tissue from histological whole-slide images (WSIs) is therefore of high clinical relevance. However, this task is challenging due to non-overlapping fragments, tissue deformation, artifacts, irregular or frayed boundaries, missing pieces, and the absence of ground-truth reassembly references. This study develops a reassembly strategy to automate and improve the manual post-operative tissue fragment reconstruction given these constraints. Results The authors present HiViTAlign, a histopathological vision transformer alignment pipeline reassembling microscopic tissue slides into a whole, which is a special variant of image registration. The task is formulated as a non-overlapping image registration problem in which adjacent fragments are identified and spatially aligned based on texture and shape cues rather than pixel overlap. The pipeline employs a three-stage coarse-to-fine vision transformer architecture to extract multi-scale features and predict translational transformation parameters between fragment pairs. To enable supervised training despite the lack of standardized datasets, a synthetic puzzle generator is introduced that produces pathology-inspired irregular fragments from masked WSI thumbnails. The model is trained in a multi-domain setting across three heterogeneous datasets: synthetic irregular fragments, regular grid tiles, and ex vivo pig organ WSIs, promoting robustness to morphological variability and staining differences. Pairwise registration achieved average mean absolute errors equivalent to <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(27.6\%-33.4\%\)</EquationSource> </InlineEquation> of fragment width with high structural preservation (NCC <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(0.96-0.97\)</EquationSource> </InlineEquation>). Global assembly using minimal spanning tree clustering yielded reconstruction errors of <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(313-378 \mu m\)</EquationSource> </InlineEquation> at <InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(1.88mm\)</EquationSource> </InlineEquation> square fragment width for typical <InlineEquation ID="IEq5"> <EquationSource Format="TEX">\(3-5\)</EquationSource> </InlineEquation> fragment puzzles, achieving <InlineEquation ID="IEq6"> <EquationSource Format="TEX">\(97\)</EquationSource> </InlineEquation>% success rate for 3-fragment cases declining due to error propagation. Completely automated HiViTAlign reassembly shows a puzzle accuracy of <InlineEquation ID="IEq7"> <EquationSource Format="TEX">\((45 \pm 22)\%\)</EquationSource> </InlineEquation> on an ex vivo pig organ tissue set. Runtime performance of <InlineEquation ID="IEq8"> <EquationSource Format="TEX">\(2.1-2.8s\)</EquationSource> </InlineEquation> per puzzle demonstrates computational feasibility for clinical integration. Conclusion Although current accuracy remains insufficient for fully automated assessment especially for larger number of fragments, the framework establishes a proof-of-concept for automated tissue reassembly and supports semi-automated workflows by proposing likely fragment adjacencies. This method supports spatial interpretation of pathological specimens and lays groundwork for future integration with 3D tissue reconstruction. The implementation of HiViTAlign and synthetic fragment generator are available open-source.

Post-operative tissue fragment puzzling using histopathological vision transformer alignment HiViTAlign

期刊论文
出版模式：

开放获取
发表日期：

2026年7月6日

查看全文

Christoph Blattgerste,
Tanzina Ferdous,
Ayk Jessen,
Maximilian Legnar,
Karl Rohr,
Claudia Scher,
Jürgen Hesser,
Cleo-Aron Weis

摘要

Background

In routine histopathology, macroscopic surgical specimens are dissected into multiple microscopic tissue fragments that are processed and analyzed independently. This fragmentation disrupts spatial continuity, making it difficult to integrate local findings into a coherent global view of tumor extent, orientation, and resection margins. Especially for malignant disease or malignoma, this is an issue of particular relevance forwhere positive margins are associated with increased mortality. Reconstructing the original macroscopic tissue from histological whole-slide images (WSIs) is therefore of high clinical relevance. However, this task is challenging due to non-overlapping fragments, tissue deformation, artifacts, irregular or frayed boundaries, missing pieces, and the absence of ground-truth reassembly references. This study develops a reassembly strategy to automate and improve the manual post-operative tissue fragment reconstruction given these constraints.

Results

The authors present HiViTAlign, a histopathological vision transformer alignment pipeline reassembling microscopic tissue slides into a whole, which is a special variant of image registration. The task is formulated as a non-overlapping image registration problem in which adjacent fragments are identified and spatially aligned based on texture and shape cues rather than pixel overlap. The pipeline employs a three-stage coarse-to-fine vision transformer architecture to extract multi-scale features and predict translational transformation parameters between fragment pairs. To enable supervised training despite the lack of standardized datasets, a synthetic puzzle generator is introduced that produces pathology-inspired irregular fragments from masked WSI thumbnails. The model is trained in a multi-domain setting across three heterogeneous datasets: synthetic irregular fragments, regular grid tiles, and ex vivo pig organ WSIs, promoting robustness to morphological variability and staining differences. Pairwise registration achieved average mean absolute errors equivalent to \(27.6\%-33.4\%\) of fragment width with high structural preservation (NCC \(0.96-0.97\) ). Global assembly using minimal spanning tree clustering yielded reconstruction errors of \(313-378 \mu m\) at \(1.88mm\) square fragment width for typical \(3-5\) fragment puzzles, achieving \(97\) % success rate for 3-fragment cases declining due to error propagation. Completely automated HiViTAlign reassembly shows a puzzle accuracy of \((45 \pm 22)\%\) on an ex vivo pig organ tissue set. Runtime performance of \(2.1-2.8s\) per puzzle demonstrates computational feasibility for clinical integration.

Conclusion

Although current accuracy remains insufficient for fully automated assessment especially for larger number of fragments, the framework establishes a proof-of-concept for automated tissue reassembly and supports semi-automated workflows by proposing likely fragment adjacencies. This method supports spatial interpretation of pathological specimens and lays groundwork for future integration with 3D tissue reconstruction. The implementation of HiViTAlign and synthetic fragment generator are available open-source.