Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Are nuclear masks all you need for improved out-of-domain generalisation? A closer look at cancer classification in histopathology

Dhananjay Tomar, Alexander Binder, Andreas Kleppe

TL;DR

This work proposes a simple approach to improve OOD generalisation for cancer detection by focusing on nuclear morphology and organisation, as these are domain-invariant features critical in cancer detection.

Abstract

Domain generalisation in computational histopathology is challenging because the images are substantially affected by differences among hospitals due to factors like fixation and staining of tissue and imaging equipment. We hypothesise that focusing on nuclei can improve the out-of-domain (OOD) generalisation in cancer detection. We propose a simple approach to improve OOD generalisation for cancer detection by focusing on nuclear morphology and organisation, as these are domain-invariant features critical in cancer detection. Our approach integrates original images with nuclear segmentation masks during training, encouraging the model to prioritise nuclei and their spatial arrangement. Going beyond mere data augmentation, we introduce a regularisation technique that aligns the representations of masks and original images. We show, using multiple datasets, that our method improves OOD generalisation and also leads to increased robustness to image corruptions and adversarial attacks. The source code is available at https://github.com/undercutspiky/SFL/

Are nuclear masks all you need for improved out-of-domain generalisation? A closer look at cancer classification in histopathology

TL;DR

This work proposes a simple approach to improve OOD generalisation for cancer detection by focusing on nuclear morphology and organisation, as these are domain-invariant features critical in cancer detection.

Abstract

Domain generalisation in computational histopathology is challenging because the images are substantially affected by differences among hospitals due to factors like fixation and staining of tissue and imaging equipment. We hypothesise that focusing on nuclei can improve the out-of-domain (OOD) generalisation in cancer detection. We propose a simple approach to improve OOD generalisation for cancer detection by focusing on nuclear morphology and organisation, as these are domain-invariant features critical in cancer detection. Our approach integrates original images with nuclear segmentation masks during training, encouraging the model to prioritise nuclei and their spatial arrangement. Going beyond mere data augmentation, we introduce a regularisation technique that aligns the representations of masks and original images. We show, using multiple datasets, that our method improves OOD generalisation and also leads to increased robustness to image corruptions and adversarial attacks. The source code is available at https://github.com/undercutspiky/SFL/

Paper Structure

This paper contains 27 sections, 2 equations, 7 figures, 31 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Proposed Method
  4. Experiments
  5. Datasets
  6. Dataset preparation
  7. Experiment setup
  8. Results on CAMELYON17 (lymph node sections)
  9. Results on BCSS (primary breast cancer)
  10. Results on Ocelot (primary non-breast cancer)
  11. Ablation Study and Discussion
  12. Impact of data augmentation
  13. Impact of $\ell_2$-regularisation
  14. Impact of mask-times-input-augmentation
  15. Impact of learned features without nuclear-mask-like features
  16. ...and 12 more sections

Figures (7)

  • Figure 1: We pass the input image (or, with 0.5 probability, input image multiplied with its nuclear segmentation mask) and its nuclear segmentation mask through the network and minimise the Binary Cross-Entropy (BCE) loss for both the input image and its mask. Additionally, we minimise the $\ell_2$-distance between the input image's embedding vector and the mask's embedding vector just before the Global Average Pooling (GAP) layer. The embedding vector is ResNet-50's penultimate layer's feature map, i.e., stage 4's last feature map.
  • Figure 2: Exemplary image ablations used in this study.
  • Figure 3: Robustness to added noise described in hendrycks2019robustness.
  • Figure 4: (a) PGD attack on models. (b) Cross-model PGD attacks where adversarial images are generated using a model from a method but the accuracy for those images is tested on models from other methods. Results are for the validation subset of each centre in CAMELYON17.
  • Figure 5: Exemplary image corruptions from hendrycks2019robustness applied to an input image used in this study.
  • ...and 2 more figures