Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Switched auxiliary loss for robust training of transformer models for histopathological image segmentation

Mustaffa Hussain, Saharsh Barve

TL;DR

The study addresses robust training of transformer models for dense histopathological segmentation of Functional Tissue Units (FTUs) across five organs using the HuBMAP+HPA data. It introduces a switched auxiliary loss to mitigate vanishing gradients in deep transformers and evaluates CoAt, PVTv2, SegFormer (with CNN baselines) on FTU segmentation, reporting Dice scores up to 0.793 on public and 0.778 on private data. The results demonstrate improved generalization and robustness of transformer-based approaches for dense medical image segmentation, supporting their broader adoption in histopathology tasks. The work also emphasizes preprocessing steps like stain normalization and domain adaptation through cross-dataset evaluation to handle real-world variability.

Abstract

Functional tissue Units (FTUs) are cell population neighborhoods local to a particular organ performing its main function.The FTUs provide crucial information to the pathologist in understanding the disease affecting a particular organ by providing information at the cellular level.In our research, we have developed a model to segment multi-organ FTUs across 5 organs namely: the kidney, large intestine, lung, prostate and spleen by utilizing the 'HuBMAP + HPA - Hacking the Human Body' competition dataset.We propose adding switched auxiliary loss for training models like the transformers to overcome the diminishing gradient problem which poses a challenge towards optimal training of deep models.Overall, our model achieved a dice score of 0.793 on the public dataset and 0.778 on the private dataset.The results supports the robustness of the proposed training methodology.The findings also bolster the use of transformers models for dense prediction tasks in the field of medical image analysis.The study assists in understanding the relationships between cell and tissue organization thereby providing a useful medium to look at the impact of cellular functions on human health.

Switched auxiliary loss for robust training of transformer models for histopathological image segmentation

TL;DR

The study addresses robust training of transformer models for dense histopathological segmentation of Functional Tissue Units (FTUs) across five organs using the HuBMAP+HPA data. It introduces a switched auxiliary loss to mitigate vanishing gradients in deep transformers and evaluates CoAt, PVTv2, SegFormer (with CNN baselines) on FTU segmentation, reporting Dice scores up to 0.793 on public and 0.778 on private data. The results demonstrate improved generalization and robustness of transformer-based approaches for dense medical image segmentation, supporting their broader adoption in histopathology tasks. The work also emphasizes preprocessing steps like stain normalization and domain adaptation through cross-dataset evaluation to handle real-world variability.

Abstract

Functional tissue Units (FTUs) are cell population neighborhoods local to a particular organ performing its main function.The FTUs provide crucial information to the pathologist in understanding the disease affecting a particular organ by providing information at the cellular level.In our research, we have developed a model to segment multi-organ FTUs across 5 organs namely: the kidney, large intestine, lung, prostate and spleen by utilizing the 'HuBMAP + HPA - Hacking the Human Body' competition dataset.We propose adding switched auxiliary loss for training models like the transformers to overcome the diminishing gradient problem which poses a challenge towards optimal training of deep models.Overall, our model achieved a dice score of 0.793 on the public dataset and 0.778 on the private dataset.The results supports the robustness of the proposed training methodology.The findings also bolster the use of transformers models for dense prediction tasks in the field of medical image analysis.The study assists in understanding the relationships between cell and tissue organization thereby providing a useful medium to look at the impact of cellular functions on human health.
Paper Structure (15 sections, 4 equations, 8 figures, 2 tables)

This paper contains 15 sections, 4 equations, 8 figures, 2 tables.

Table of Contents

  1. 1. Introduction
  2. 2. Literature Review
  3. 3. Exploratory Dataset Analysis (EDA)
  4. 4. Proposed Methodology
  5. 4.1 Models
  6. 4.1.1 CoAt (Convolution Attention)
  7. 4.1.2 Pyramid Vision Transformer-v2 (PVTv2)
  8. 4.1.3 Segformer
  9. 4.2 Metric
  10. 4.3 Losses
  11. 5. Training Pipeline
  12. 6. Inference Pipeline
  13. Results
  14. Conclusion
  15. Appendix

Figures (8)

  • Figure 1: Representation of Organ-Specific Overlays from the Training Dataset
  • Figure 2: The distribution of organ classes within the training dataset
  • Figure 3: Model training Framework
  • Figure 4: Inference Framework
  • Figure 5: Sample HuBMAP test image with and without stain normalization
  • ...and 3 more figures