DCMIL: A Progressive Representation Learning of Whole Slide Images for Cancer Prognosis Analysis

TL;DR

DCMIL presents a novel two-curriculum framework for prognostic learning from gigapixel whole-slide images without dense annotations. Curriculum I uses saliency-guided, cross-scale instance encoding to leverage multi-magnification information, while Curriculum II performs adaptive soft-bag prognosis inference with constrained self-attention and triple-tier contrastive learning to maximize intra- and inter-bag discrimination. Across 12 TCGA cancer types and nearly 6,000 patients, DCMIL achieves state-of-the-art concordance indices and delivers interpretable saliency maps and representative soft-bag embeddings that align with histopathology. By quantifying tumor heterogeneity and enabling prognosis prediction directly from WSIs, DCMIL offers a scalable, interpretable tool for clinical decision support and potential biomarker discovery.

Abstract

The burgeoning discipline of computational pathology shows promise in harnessing whole slide images (WSIs) to quantify morphological heterogeneity and develop objective prognostic modes for human cancers. However, progress is impeded by the computational bottleneck of gigapixel-size inputs and the scarcity of dense manual annotations. Current methods often overlook fine-grained information across multi-magnification WSIs and variations in tumor microenvironments. Here, we propose an easy-to-hard progressive representation learning, termed dual-curriculum contrastive multi-instance learning (DCMIL), to efficiently process WSIs for cancer prognosis. The model does not rely on dense annotations and enables the direct transformation of gigapixel-size WSIs into outcome predictions. Extensive experiments on twelve cancer types (5,954 patients, 12.54 million tiles) demonstrate that DCMIL outperforms standard WSI-based prognostic models. Additionally, DCMIL identifies fine-grained prognosis-salient regions, provides robust instance uncertainty estimation, and captures morphological differences between normal and tumor tissues, with the potential to generate new biological insights. All codes have been made publicly accessible at https://github.com/tuuuc/DCMIL.

Figures (6)

• Figure 1: Overview of the framework.$\textbf{a}$, The overall workflow, including cancer resection and staining, digital slide scanning, DCMIL construction, and model validation. $\textbf{b}$, The pipeline of DCMIL comprises two easy-to-hard curriculums, including 1) saliency-guided weakly-supervised instance encoding with cross-scale tiles, and 2) contrastive-enhanced soft-bag prognosis inference.
• Figure 2: The detailed architectures on Curriculum I. For the $s$-th branch, its architecture mainly consists of a linear layer $\mathbb{P}^s$, a transformer encoder $\mathbb{F}^s$, an attention aggregator $\mathbb{G}^s$ and a multilayer perceptron (MLP) head $\mathbb{C}^s$.
• Figure 3: The detailed architectures on Curriculum II, containing a soft-bag learning module $\mathbb{B}$, a constrained self-attention module $\mathbb{D}$, and a contrastive-enhanced prognosis inference function $\mathbb{S}$.
• Figure 4: The KM curves along with p-values on twelve datasets.
• Figure 5: Visualizations of saliency maps across multi-magnification WSIs on Curriculum I for COAD dataset.