Cross-Stream Interactions: Segmentation of Lung Adenocarcinoma Growth Patterns

• Published in: Computational Mathematics Modeling in Cancer Analysis Vol. 13574; pp. 78 - 90
• Main Authors: Pan, Xiaoxi, Zhang, Hanyun, Grapa, Anca-Ioana, AbdulJabbar, Khalid, Raza, Shan E Ahmed, Cheung, Ho Kwan Alvin, Karasaki, Takahiro, Quesne, John Le, Moore, David A., Swanton, Charles, Yuan, Yinyin
• Format: Book Chapter
• Language: English
• Published: Switzerland Springer 2022; Springer Nature Switzerland
• Series: Lecture Notes in Computer Science
• Subjects: Growth patterns; Histology; Semantic segmentation
• ISBN: 9783031172656; 3031172655
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-031-17266-3_8

Lung adenocarcinoma has histologically distinct growth patterns that have been associated with patient prognosis. Precision segmentation of growth patterns in routine histology samples is challenging due to the complexity of patterns and high intra-class variability. In this paper, we present a novel model with a multi-stream architecture, Cross-Stream Interactions (CroSIn), which fully considers crucial interactions across scales to gather abundant information. The first-order attention introduces contextual information at an early stage to guide low-level feature encoding. The second-order attention then focuses on learning high-level feature relations among scales to extract discriminative features. Experimental results show interactions at both low- and high-level feature learning stages are crucial in performance improvement. The proposed method outperforms state-of-the-art networks, achieving an average Dice of 60.34% $$60.34\%$$ at patch level, and an average accuracy of 65.31% $$65.31\%$$ at sample level, which is also verified in an independent cohort.