AttentioScope: a multi-indicator quantification framework with attention mechanism based on thyroid fine-needle aspiration smear images
Year: 2026
Type: article
Abstract: The interpretability of deep learning models is crucial for their clinical application in computational pathology. Although attention-based methods can provide intuitive heatmap explanations, their reliability on challenging samples still requires systematic, fine-grained evaluation tools to assess model performance more comprehensively. To this end, this paper proposes a multi-metric quantitative framework based on thyroid fine-needle aspiration smear images to evaluate the comprehensive performance of different models through visual analysis. First, the framework assesses the reliability and focus capability of attention distributions by computing multi-dimensional quantitative metrics such as entropy, concentration, stability, and compactness of the attention heatmaps. Second, by correlating the heatmaps with underlying image features, it identifies whether models rely on pathology-irrelevant low-level features. To validate the effectiveness of the framework, we systematically evaluated four mainstream attention-based multiple instance learning (MIL) models (CLAM, TransMIL, PEGTB-MIL, ADMIL) on athyroid fine-needle aspiration smear image dataset. Experimental results show that in challenging samples of high-performance models, attention mechanisms may exhibit patterns such as feature hijacking or spatial bias, which traditional performance metrics fail to effectively reveal. This framework aims to conduct a fine-grained evaluation of attention-based models, assisting in identifying potential hidden deficiencies in challenging samples, thereby providing a reference framework for model assessment and optimization.
Source: Frontiers in Endocrinology
Authors: Qinlong Lv, En Fan, Zonghua Wen, Yunming Qiu, Yajun Hu
Institutions: Shaoxing University, Shenzhen University
Language: English
Topic: AI in cancer detection
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $3,801
OpenAlex ID: W7226005771
