Year: 2025
Type: dissertation
Abstract: This study aims to leverage multimodal deep learning techniques to predict the efficacy of salvage chemoradiotherapy for patients with regional lymph node recurrence following curative resection of esophageal squamous cell carcinoma (ESCC).The challenge of accurately predicting treatment efficacy in this context is compounded by the limited availability of usable data, alongside a heavy reliance on manual feature selection processes that may overlook crucial information.Given these constraints, integrating multimodal information-such as medical images, clinical texts, and tabular datasets-presents a promising approach to enhance the predictive performance of treatment outcomes.Despite the potential advantages, there exists a notable research gap addressing this specific issue through the comprehensive integration of multimodal data.Prior studies often focus on single-modality approaches, thereby failing to capture the richness of information inherent in diverse data sources.To address this gap, we conduct a retrospective study that meticulously tracks the clinical information of patients who undergo chemoradiotherapy after experiencing lymph node recurrence.This involves collecting a robust multimodal dataset that encompasses various types of data.Building upon this dataset, we design a sophisticated multimodal architecture that incorporates three distinct networks, each dedicated to processing medical images, textual information, and tabular data concurrently.This architecture employs advanced attention mechanisms to dynamically weigh the importance of different features, coupled with Fourier transform dual paths that facilitate seamless feature iv fusion.The innovative design not only enhances the model's ability to extract relevant features but also ensures that interactions between different modalities are effectively captured.Through a rigorous comparative analysis of our proposed model against traditional single-modality approaches, we observe a significant enhancement in performance metrics.Specifically, the model achieves notable average precision, recall, and F1score values of 91.96%, 91.18%, and 90.86%, respectively.Furthermore, the model reaches peak values of 97.20%, 97.10%, and 97.00% for these metrics, highlighting the efficacy of integrating multimodal data.The findings from this study underscore the value of incorporating diverse data sources in clinical decision-making, demonstrating that the integration of image, text, and tabular data can substantially enrich the information available for predicting treatment efficacy.Ultimately, this research fills a critical gap in the existing literature and presents an exciting avenue for future research and development.
DOI: 10.15099/0002001368
Author: 張, 晗
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W6942654488
Cites:
Cited by:
Related to: