Enhancing poorly differentiated lung cancer classification with rotary position embedding and sparse attention in multiple instance learning

In clinical practice, diagnosing poorly differentiated non-small cell lung cancer (NSCLC) typically requires immunohistochemistry (IHC) to accurately distinguish between cancer subtypes. The high cost and time-consuming nature of this process significantly limit its clinical applicability. Furthermo...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:Biomedical signal processing and control Jg. 113; S. 108699
Hauptverfasser: Chu, Hongbo, Chang, Zhengsong, Shao, Zihan, Deng, Pengyun, Wang, Yan, Guo, Xiaojing, Wang, Yang, Yang, Jiao, Guan, Tian, Li, Xiaomei
Format: Journal Article
Sprache:Englisch
Veröffentlicht: Elsevier Ltd 01.03.2026
Schlagworte:
Deep learning
Histopathological classification
Multiple instance learning
Poorly differentiated lung cancer
Transformer
Deep learning
Transformer
Histopathological classification
Poorly differentiated lung cancer
Multiple instance learning
ISSN:1746-8094
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:In clinical practice, diagnosing poorly differentiated non-small cell lung cancer (NSCLC) typically requires immunohistochemistry (IHC) to accurately distinguish between cancer subtypes. The high cost and time-consuming nature of this process significantly limit its clinical applicability. Furthermore, the limited availability of tissue samples from advanced-stage patients poses a significant challenge for subsequent diagnosis. Consequently, leveraging novel AI technologies is critical to overcoming this challenge. Hence, we propose an innovative Multi-Instance Learning (MIL) algorithm, RoSA-MIL. This algorithm is based on a novel MIL framework that integrates relative position encoding and sparse attention mechanisms. Relative position encoding enhances the model’s comprehension of spatial relationships between samples, thereby improving its ability to capture local context. The sparse attention mechanism reduces computational overhead by eliminating redundant calculations, thereby enhancing both efficiency and scalability. Experiments conducted on diverse datasets from three independent centers demonstrate that the RoSA-MIL model outperforms existing methods, exhibiting significant improvements in diagnostic accuracy, robustness, and efficiency when processing large-scale data.
ISSN:1746-8094
DOI:10.1016/j.bspc.2025.108699