A novel method for unsupervised scanner-invariance with DCAE model

Andrew Moyes, Kun Zhang, Liping Wang, Ji Ming, Daniel Crookes, Huiyu ZhouLeicester 2018

Abstract

Automated analysis of histopathology whole-slide images is impeded by the scanner-dependent variance introduced in the slide scanning process. This work presents a novel dual-channel auto-encoder based model with a multi-component loss which learns a scanner-invariant representation of histopathology images. The learned representation can be used for a number of histopathology-related applications where images are captured from different scanners such as nuclei detection and cancer segmentation. The approach is validated on a set of lung tissue sub-images extracted from whole slide images. This method achieves a 50% improvement in SSIM score on tissue masks derived from the learned representation compared to related methods. To the best of the author’s knowledge, this is the first work which explicitly learns a scanner-invariant representation of histopathology images from multiple domains simultaneously without labelled data or expensive preprocessing techniques.

Introduction

In computational pathology, the "gold standard" for diagnosis is the Whole-Slide Image (WSI). However, these digital slides suffer from a major consistency problem: scanner-dependent variance. Just as two different smartphone cameras might produce different colors and sharpness for the same scene, different medical scanners (like those from Hamamatsu, Aperio, or Philips) introduce unique discrepancies in illumination, contrast, and resolution.

For a human pathologist, these differences are minor. For an AI model, they are catastrophic. A model trained to detect cancer cells on images from Scanner A often fails when deployed on images from Scanner B. Existing solutions usually involve "stain normalization," but these often require a target reference image and can be computationally expensive or lose critical structural detail.

Methods: The Dual-Channel Auto-Encoder (DCAE)

Instead of trying to change the pixels of the image, the researchers propose changing the representation of the image in the computer's latent space. The core of this approach is the Dual-Channel Auto-Encoder (DCAE).

The proposed DCAE workflow, from whole-slide registration to scanner-invariant feature representation.

Architecture and Loss Functions

Two identical auto-encoders are trained in tandem. One processes the "Source" scanner images, and the other processes the "Target" scanner images of the same tissue. The learning process is guided by a multi-component loss function:

  1. Reconstruction Loss ($L_r$): Ensures the model understands the image by forcing it to rebuild the original input.
  2. Feature Loss ($L_f$): Forces the internal feature vectors of both channels to be as similar as possible for the same tissue sample.
  3. Cluster Loss ($L_c$): A modified "center loss" that uses K-Means to create "pseudo-labels." This forces the model to group similar tissue types together in the feature space, even without human-provided labels.

To prevent any single loss from dominating, the authors use a geometric approach to balance these components, effectively maintaining an "equilateral triangle" relationship between the three loss types during training.

Experimental Setup

The study utilized a custom dataset of 60 whole-slide images representing 30 different lung tissue specimens. Each specimen was scanned twice: once using an Aperio scanner and once using a Philips scanner. The WSIs were split into 263 sub-image pairs, from which approximately 300,000 finely aligned patch pairs were extracted for training using ORB feature matching.

Key Results and Insights

The DCAE model significantly outperformed existing techniques in creating scanner-invariant masks. Quantitatively, the model achieved a 50% improvement in Structural Similarity Index (SSIM) scores compared to related methods like StaNoSA.

Qualitative comparison of tissue masks and feature representations across different loss configurations.

Qualitative results (as seen in the figure above) show that the DCAE-FC (Feature + Cluster) model preserves fine details, such as the boundaries of nuclei, which are often blurred or lost in standard RGB-pixel or sparse auto-encoder methods. PCA visualizations confirmed that the combined feature-cluster loss created much clearer separation between different tissue types in the latent space compared to using feature loss alone.

Conclusion and Future Work

By using two separate auto-encoders, each network can specialize in the specific "quirks" of its assigned scanner brand, allowing for a more detailed representation. While the current architecture requires a dedicated channel for each scanner type—a potential scalability hurdle—it proves that unsupervised learning can effectively bridge the gap between hardware domains.

Future work aims to develop representations invariant to both scanner and staining variations simultaneously and to test the model on larger, public datasets like MITOS-ATYPIA to prove broader clinical utility.

Summary updated

Abstract

Automated analysis of histopathology whole-slide images is impeded by the scannerdependent
\nvariance introduced in the slide scanning process. This work presents a novel
\ndual-channel auto-encoder based model with a multi-component loss which learns a
\nscanner-invariant representation of histopathology images. The learned representation
\ncan be used for a number of histopathology-related applications where images are captured
\nfrom different scanners such as nuclei detection and cancer segmentation. The
\napproach is validated on a set of lung tissue sub-images extracted from whole slide images.
\nThis method achieves a 50% improvement in SSIM score on tissue masks derived
\nfrom the learned representation compared to related methods. To the best of the author’s
\nknowledge, this is the first work which explicitly learns a scanner-invariant representation
\nof histopathology images from multiple domains simultaneously without labelled
\ndata or expensive preprocessing techniques.
BibTeX
@inproceedings{moyes2018novel,
  author = {Andrew Moyes and Kun Zhang and Liping Wang and Ji Ming and Daniel Crookes and Huiyu Zhou},
  title = {{A novel method for unsupervised scanner-invariance with DCAE model}},
  booktitle = {Leicester},
  year = {2018},
}
Figure