Linear Attention-based Multiple Instance Learning for Computational Pathology

Charlotte Richter, Daniel Reisenbüchler, Nadine S. Schaadt, Friedrich Feuerhake, Dorit Merhof

Keywords: Computational Pathology, Multiple Instance Learning, Linear Attention, Whole Slide Image Analysis

TL;DR: A novel multiple instance learning approach with linear scale for Whole Slide Image Classification

Abstract: Deep learning–based analysis of gigapixel whole slide images (WSIs) in computational pathology (CPath) typically relies on patch-level feature extraction and instance aggregation, with attention-based contextualization at the core of state-of-the-art methods. However, scalability is a major challenge due to the vast number of patches. Therefore, we introduce linear attention based multiple-instance learning (Lin-MIL), which transposes and interchanges the calculations of queries, keys, and values in the attention mechanism. By leveraging linear attention, Lin-MIL reduces computational complexity from $\mathcal{O}(n^2 d)$ to $\mathcal{O}(n d^2)$, compared to vanilla self-attention. Despite this efficiency gain, Lin-MIL outperforms 12 baseline methods across biomarker, mutation, and tumor classification benchmarks, while also demonstrating robust out-of-domain performance. Moreover, its qualitative attention maps highlight diagnostically relevant regions. In summary, Lin-MIL provides increased performance as well as enhanced scalability and interpretability for a range of computational pathology tasks. Code available at https://github.com/charlotterchtr/Lin-MIL

Submission Number: 10