[Paper Review] Reinforced Auto-Zoom Net: Towards Accurate and Fast Breast Cancer Segmentation in Whole-slide Images

Nanqing Dong, Michael Kampffmeyer+4 more|arXiv (Cornell University)|Jul 29, 2018

AI in cancer detection12 references9 citations

TL;DR

This paper proposes Reinforced Auto-Zoom Net (RAZN), a deep reinforcement learning framework that selectively zooms into suspicious regions of whole-slide images (WSIs) for accurate and efficient breast cancer segmentation. By learning a policy to decide when to zoom in, RAZN improves segmentation accuracy on imbalanced, noisy data while reducing inference time compared to single- and multi-scale baselines, achieving a mean IOU of 0.49 and relative inference time of 2.71 ± 0.57.

ABSTRACT

Convolutional neural networks have led to significant breakthroughs in the domain of medical image analysis. However, the task of breast cancer segmentation in whole-slide images (WSIs) is still underexplored. WSIs are large histopathological images with extremely high resolution. Constrained by the hardware and field of view, using high-magnification patches can slow down the inference process and using low-magnification patches can cause the loss of information. In this paper, we aim to achieve two seemingly conflicting goals for breast cancer segmentation: accurate and fast prediction. We propose a simple yet efficient framework Reinforced Auto-Zoom Net (RAZN) to tackle this task. Motivated by the zoom-in operation of a pathologist using a digital microscope, RAZN learns a policy network to decide whether zooming is required in a given region of interest. Because the zoom-in action is selective, RAZN is robust to unbalanced and noisy ground truth labels and can efficiently reduce overfitting. We evaluate our method on a public breast cancer dataset. RAZN outperforms both single-scale and multi-scale baseline approaches, achieving better accuracy at low inference cost.

Motivation & Objective

• To address the challenge of accurate and fast breast cancer segmentation in high-resolution whole-slide images (WSIs), which are constrained by hardware and field of view limitations.
• To overcome data imbalance and noise in pixel-level annotations, which often lead to overfitting in conventional deep learning approaches.
• To develop a framework that mimics the pathologist’s selective zoom-in behavior using reinforcement learning for improved efficiency and robustness.
• To evaluate the effectiveness of reinforcement learning in medical image segmentation compared to traditional multi-scale training strategies.

Proposed method

• RAZN employs a policy network trained via policy gradient methods to decide whether to zoom in on a given region of interest in a WSI patch.
• The framework uses two resolution scales: coarse (64×64) and fine (128×128), both resized to 256×256 for input to the segmentation network.
• The segmentation network is based on ResNet18, with modified residual blocks to reduce computational cost in the policy network.
• The policy network is trained end-to-end with the segmentation network using the Adam optimizer and a step-wise learning rate decay.
• Zoom-in decisions are stochastic, allowing the model to avoid unnecessary upsampling and training on uninformative regions.
• The model uses mean IOU and weighted IOU as evaluation metrics, with relative inference time normalized to a single-scale baseline.

Experimental results

Research questions

• RQ1 Can a reinforcement learning-based zoom-in policy improve segmentation accuracy on imbalanced and noisy whole-slide images?
• RQ2 Does selective zooming reduce overfitting on normal and benign tissue while enhancing feature learning on carcinoma regions?
• RQ3 How does the inference speed of a reinforcement learning-based auto-zoom mechanism compare to fixed multi-scale approaches?
• RQ4 Can the proposed framework outperform both single-scale and multi-scale baselines in terms of accuracy and efficiency?

Key findings

• RAZN achieved a mean IOU of 0.49, significantly outperforming the single-scale baseline (0.38) and the multi-scale MS model (0.18).
• The weighted IOU for RAZN was 0.11, the highest among all methods, indicating improved performance on rare or imbalanced classes.
• RAZN achieved a relative inference time of 2.71 ± 0.57, which is faster than both multi-scale baselines (MS: 5.06, Attention: 5.16), demonstrating efficiency.
• The MS model performed poorly (mean IOU: 0.18) due to data augmentation exacerbating class imbalance and introducing noise.
• The Attention model showed moderate improvement but was still outperformed by RAZN, suggesting that scale fusion may not be optimal for small, fixed-size cancer cells.
• RAZN’s stochastic zoom-in policy reduced overfitting by avoiding unnecessary training on non-suspicious regions, improving robustness to noisy labels.