Applications of Multiple Instance Learning in Computational Pathology: A Survey

The Hong Kong University of Science and Technology (Guangzhou)

数据科学与分析学域

PhD Qualifying Examination

By Mr. GUO Pengyu

摘要

This survey reviews annotation-efficient strategies in computational pathology, with a focus on multiple-instance learning (MIL). We first outline the high cost of pixel-level labeling and summarize semi-supervised and weakly supervised alternatives. We then dissect the MIL pipeline—comparing explicit versus attention-based instance selection, detailing pooling and Transformer-based aggregation schemes, and formalizing three core down stream tasks (classification, weakly supervised segmentation, and continuous-value regression). Next, we examine how MIL enables multimodal fusion via both modality-alignment and modality-aggregation techniques that integrate histology with text, genomics, and clinical data. We then catalog key public datasets, highlight their slide-level annotation bias, and review standard evaluation metrics. Finally, we identify open challenges and sketch future directions, for example the adoption of multi-modal data. This roadmap aims to guide researchers toward robust, interpretable, and scalable MIL solutions for real-world pathology.

PQE Committee

Chair of Committee: Prof. TANG Nan

Prime Supervisor: Prof. CHEN Lei

Co-Supervisor: Prof. ZHANG Yongqi

Examiner: Prof. LI Lei