Cross-attention-based saliency inference for predicting cancer metastasis on whole slide images

TL;DR

This paper introduces CASiiMIL, a novel cross-attention-based MIL method that improves detection of small tumor metastases in whole slide images without annotations, enhancing sensitivity, interpretability, and cross-center generalizability.

Contribution

The paper proposes a new saliency-informed attention mechanism and negative representation learning for MIL, significantly improving early tumor detection in WSIs without requiring annotations.

Findings

Outperforms state-of-the-art MIL methods on metastasis detection datasets.

Demonstrates high sensitivity for small tumor lesions.

Shows strong cross-center generalizability and interpretability.

Abstract

Although multiple instance learning (MIL) methods are widely used for automatic tumor detection on whole slide images (WSI), they suffer from the extreme class imbalance within the small tumor WSIs. This occurs when the tumor comprises only a few isolated cells. For early detection, it is of utmost importance that MIL algorithms can identify small tumors, even when they are less than 1% of the size of the WSI. Existing studies have attempted to address this issue using attention-based architectures and instance selection-based methodologies, but have not yielded significant improvements. This paper proposes cross-attention-based salient instance inference MIL (CASiiMIL), which involves a novel saliency-informed attention mechanism, to identify breast cancer lymph node micro-metastasis on WSIs without the need for any annotations. Apart from this new attention mechanism, we introduce a negative representation learning algorithm to facilitate the learning of saliency-informed attention weights for improved sensitivity on tumor WSIs. The proposed model outperforms the state-of-the-art MIL methods on two popular tumor metastasis detection datasets, and demonstrates great cross-center generalizability. In addition, it exhibits excellent accuracy in classifying WSIs with small tumor lesions. Moreover, we show that the proposed model has excellent interpretability attributed to the saliency-informed attention weights. We strongly believe that the proposed method will pave the way for training algorithms for early tumor detection on large datasets where acquiring fine-grained annotations is practically impossible.