Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Deep Learning-Based Fixation Type Prediction for Quality Assurance in Digital Pathology

Oskar Thaeter, Tanja Niedermair, Johannes Raffler, Ralf Huss, Peter J. Schüffler

TL;DR

The paper tackles mislabelling of fixation type in digital pathology and proposes a fast, thumbnail-based deep learning approach to verify fixation type before high-resolution scanning. Using four pathology-pretrained vision transformers and both whole-slide and tiled-slide strategies, the best model (UNI with soft voting) attains an AUROC of $0.88$ on TCGA and processes each slide in $21\ \mathrm{ms}$, enabling high-throughput pre-scan QC. External evaluation reveals a notable domain shift across non-Leica scanners, with AUROCs dropping to $0.72$, highlighting cross-scanner generalization as the primary remaining challenge. The work demonstrates the practicality of pre-scan thumbnail analysis for quality control in digital pathology and points to extensions to additional slide-level annotations and modalities, as well as domain-adaptation techniques to improve robustness.

Abstract

Accurate annotation of fixation type is a critical step in slide preparation for pathology laboratories. However, this manual process is prone to errors, impacting downstream analyses and diagnostic accuracy. Existing methods for verifying formalin-fixed, paraffin-embedded (FFPE), and frozen section (FS) fixation types typically require full-resolution whole-slide images (WSIs), limiting scalability for high-throughput quality control. We propose a deep-learning model to predict fixation types using low-resolution, pre-scan thumbnail images. The model was trained on WSIs from the TUM Institute of Pathology (n=1,200, Leica GT450DX) and evaluated on a class-balanced subset of The Cancer Genome Atlas dataset (TCGA, n=8,800, Leica AT2), as well as on class-balanced datasets from Augsburg (n=695 [392 FFPE, 303 FS], Philips UFS) and Regensburg (n=202, 3DHISTECH P1000). Our model achieves an AUROC of 0.88 on TCGA, outperforming comparable pre-scan methods by 4.8%. It also achieves AUROCs of 0.72 on Regensburg and Augsburg slides, underscoring challenges related to scanner-induced domain shifts. Furthermore, the model processes each slide in 21 ms, $400\times$ faster than existing high-magnification, full-resolution methods, enabling rapid, high-throughput processing. This approach provides an efficient solution for detecting labelling errors without relying on high-magnification scans, offering a valuable tool for quality control in high-throughput pathology workflows. Future work will improve and evaluate the model's generalisation to additional scanner types. Our findings suggest that this method can increase accuracy and efficiency in digital pathology workflows and may be extended to other low-resolution slide annotations.

Deep Learning-Based Fixation Type Prediction for Quality Assurance in Digital Pathology

TL;DR

The paper tackles mislabelling of fixation type in digital pathology and proposes a fast, thumbnail-based deep learning approach to verify fixation type before high-resolution scanning. Using four pathology-pretrained vision transformers and both whole-slide and tiled-slide strategies, the best model (UNI with soft voting) attains an AUROC of on TCGA and processes each slide in , enabling high-throughput pre-scan QC. External evaluation reveals a notable domain shift across non-Leica scanners, with AUROCs dropping to , highlighting cross-scanner generalization as the primary remaining challenge. The work demonstrates the practicality of pre-scan thumbnail analysis for quality control in digital pathology and points to extensions to additional slide-level annotations and modalities, as well as domain-adaptation techniques to improve robustness.

Abstract

Accurate annotation of fixation type is a critical step in slide preparation for pathology laboratories. However, this manual process is prone to errors, impacting downstream analyses and diagnostic accuracy. Existing methods for verifying formalin-fixed, paraffin-embedded (FFPE), and frozen section (FS) fixation types typically require full-resolution whole-slide images (WSIs), limiting scalability for high-throughput quality control. We propose a deep-learning model to predict fixation types using low-resolution, pre-scan thumbnail images. The model was trained on WSIs from the TUM Institute of Pathology (n=1,200, Leica GT450DX) and evaluated on a class-balanced subset of The Cancer Genome Atlas dataset (TCGA, n=8,800, Leica AT2), as well as on class-balanced datasets from Augsburg (n=695 [392 FFPE, 303 FS], Philips UFS) and Regensburg (n=202, 3DHISTECH P1000). Our model achieves an AUROC of 0.88 on TCGA, outperforming comparable pre-scan methods by 4.8%. It also achieves AUROCs of 0.72 on Regensburg and Augsburg slides, underscoring challenges related to scanner-induced domain shifts. Furthermore, the model processes each slide in 21 ms, faster than existing high-magnification, full-resolution methods, enabling rapid, high-throughput processing. This approach provides an efficient solution for detecting labelling errors without relying on high-magnification scans, offering a valuable tool for quality control in high-throughput pathology workflows. Future work will improve and evaluate the model's generalisation to additional scanner types. Our findings suggest that this method can increase accuracy and efficiency in digital pathology workflows and may be extended to other low-resolution slide annotations.
Paper Structure (18 sections, 1 equation, 6 figures, 7 tables)

This paper contains 18 sections, 1 equation, 6 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methods
  4. Datasets
  5. Pre-Processing
  6. Vision Backbones
  7. Classification Approaches
  8. Whole-Slide Classification
  9. Tiled-Slide Classification
  10. Classification Head
  11. Results
  12. Model Selection
  13. External Evaluation
  14. Discussion
  15. Comparison to Prior Work
  16. ...and 3 more sections

Figures (6)

  • Figure 1: Example slides of FFPE and FS fixation types shown as thumbnails and at high magnification.
  • Figure 2: Possible applications of our approach in a digital pathology workflow. The fixation-type annotation can be verified before the high-resolution scan is conducted, allowing for re-annotation, re-examination, or skipping the scan entirely. Our approach can also be used as an efficient quality control measure in digital pathology archives. Created with BioRender.com
  • Figure 3: Tissue processing workflow and the two fixation methods used in this study. (a) Tissue samples are fixed, sectioned, mounted, and stained. (b) In FFPE processing, tissue is formalin-fixed and paraffin-embedded before microtome sectioning. (c) In FS processing, tissue is snap-frozen and sectioned in a cryostat. Created with BioRender.com
  • Figure 4: Stretched slide thumbnail with tile grids overlaid. Red: L grid ($4 \times 8$), green: M grid ($2 \times 4$), blue: S grid ($1 \times 2$).
  • Figure 5: Pre-processing workflow for the M configuration. The thumbnail is stretched to $896 \times 1{,}792$ px, resized to $448 \times 896$ px, and divided into a $2 \times 4$ grid of $224 \times 224$ px tiles (green).
  • ...and 1 more figures