MergeUp-augmented Semi-Weakly Supervised Learning for WSI Classification

TL;DR

This paper tackles WSI classification under weak supervision by converting to a semi-weakly supervised setting. It introduces AdaPse to adaptively assign pseudo bag labels and MergeUp to mix bags across categories, all within a student–teacher SSL framework that enforces consistency between weak (identity) and strong (MergeUp) augmentations. Empirical results on CAMELYON-16, BRACS, and TCGA-LUNG show substantial improvements over state-of-the-art MIL methods, with ablations confirming the benefits of adaptive pseudo labeling and inter-category augmentation. The approach offers a scalable, annotation-free path to improved MIL for pathology, with potential impact on diagnostic accuracy and data efficiency.

Abstract

Recent advancements in computational pathology and artificial intelligence have significantly improved whole slide image (WSI) classification. However, the gigapixel resolution of WSIs and the scarcity of manual annotations present substantial challenges. Multiple instance learning (MIL) is a promising weakly supervised learning approach for WSI classification. Recently research revealed employing pseudo bag augmentation can encourage models to learn various data, thus bolstering models' performance. While directly inheriting the parents' labels can introduce more noise by mislabeling in training. To address this issue, we translate the WSI classification task from weakly supervised learning to semi-weakly supervised learning, termed SWS-MIL, where adaptive pseudo bag augmentation (AdaPse) is employed to assign labeled and unlabeled data based on a threshold strategy. Using the "student-teacher" pattern, we introduce a feature augmentation technique, MergeUp, which merges bags with low-priority bags to enhance inter-category information, increasing training data diversity. Experimental results on the CAMELYON-16, BRACS, and TCGA-LUNG datasets demonstrate the superiority of our method over existing state-of-the-art approaches, affirming its efficacy in WSI classification.

Figures (6)

• Figure 1: Mislabeling issue faced by different pseudo bag augmentation strategies.
• Figure 2: Overview of proposed semi-weakly supervised learning framework SWS-MIL. (a) All instances in WSIs are augmented by adaptive pseudo bags for the teacher model, and MergeUp augmentation is then applied for the student model. (b) Adaptive pseudo bag augmentation determines whether the pseudo bag label is correct and recycles the incorrectly labeled instances. (c) MergeUp augmentation blends instances from low-priority pseudo bags into high-priority pseudo bags.
• Figure 3: Pseudo label accuracy $PseAcc$ of different pseudo bag augmentation methods during training.
• Figure 4: Heatmap of our method in the CAMELYON-16 dataset. (a) and (b) are macro and micro metastasis cases. In the column of "Ground Truth", cancer and non-cancer regions are delineated in red and green, respectively. The "Heatmap" column represents the prediction results of SWS-MIL, where a redder color indicates greater importance of instance.
• Figure 5: Heatmaps of a case from CAMELYON-16 generated by ABMIL, CLAM, and SWS-MIL respectively. In the "Ground Truth" column, tumor / normal regions are delineated by red / green lines. Brighter red colors indicate higher tumor probabilities.