arXiv subjects

Tu Dan

Publications and source records attributed to Tu Dan.

3 recordsLinked to original sources

arXiv2025

A survival analysis of glioma patients using topological features and locations of tumors

Tumor shape plays a critical role in influencing both growth and metastasis. We introduce a novel topological radiomic feature derived from persistent homology to characterize tumor shape, focusing on its association with time-to-event outcomes in gliomas. These features effectively capture diverse tumor shape patterns that are not represented by conventional radiomic measures. To incorporate these features into survival analysis, we employ a functional Cox regression model in which the topological features are represented in a functional space. We further include interaction terms between shape features and tumor location to capture lobe-specific effects. This approach enables interpretable assessment of how tumor morphology relates to survival risk. We evaluate the proposed method in two case studies using radiomic images of high-grade and low-grade gliomas. The findings suggest that the topological features serve as strong predictors of survival prognosis, remaining significant after adjusting for clinical variables, and provide additional clinically meaningful insights into tumor behavior.

stat.ME↗

arXiv2024

Multiomics-based Outcome Prediction for the Treatment of Brain Metastases with Personalized Ultra-fractionated Stereotactic Adaptive Radiotherapy (PULSAR)

Purpose: We aimed to develop a data-driven multiomics approach integrating radiomics, dosiomics, and delta features to predict treatment response at an earlier stage (intra-treatment) for brain metastases (BMs) patients treated with PULSAR. Methods: We conducted a retrospective study of 39 patients with 69 BMs treated with PULSAR. Radiomics, dosiomics, and delta features were extracted from pretreatment and intra-treatment MRI scans and dose distributions. Six individual models and an ensemble feature selection (EFS) model were constructed using SVM and evaluated via stratified 5-fold cross-validation. The classification task distinguished lesions with >20% volume reduction at follow-up. We assessed performance metrics including sensitivity, specificity, accuracy, precision, F1 score, and AUC. Various feature extraction and ensemble selection scenarios were explored to enhance model robustness and reduce overfitting. Results: The EFS model, integrating features from pre-treatment radiomics and dosiomics, intra-treatment radiomics, and delta-radiomics, outperformed six individual models. It achieved an AUC of 0.979, accuracy of 0.917, and F1 score of 0.821. Among the top 9 features, six were derived from post-wavelet transformation, and three from original images. The discrete wavelet transform decomposes volumetric images into multi-resolution components, offering a more comprehensive characterization of underlying structures. Conclusion: Our study demonstrated the feasibility of employing a data-driven multiomics approach to predict tumor volume changes in BMs patients with PULSAR. The EFS model demonstrates enhanced performance compared with six individual models, emphasizing the importance of integrating both pretreatment and intra-treatment data. This application holds promise for optimizing BMs management, potentially mitigating risks associated with under- or over-treatment.

physics.med-ph↗

arXiv2016

Memory Matters: Convolutional Recurrent Neural Network for Scene Text Recognition

Text recognition in natural scene is a challenging problem due to the many factors affecting text appearance. In this paper, we presents a method that directly transcribes scene text images to text without needing of sophisticated character segmentation. We leverage recent advances of deep neural networks to model the appearance of scene text images with temporal dynamics. Specifically, we integrates convolutional neural network (CNN) and recurrent neural network (RNN) which is motivated by observing the complementary modeling capabilities of the two models. The main contribution of this work is investigating how temporal memory helps in an segmentation free fashion for this specific problem. By using long short-term memory (LSTM) blocks as hidden units, our model can retain long-term memory compared with HMMs which only maintain short-term state dependences. We conduct experiments on Street View House Number dataset containing highly variable number images. The results demonstrate the superiority of the proposed method over traditional HMM based methods.

cs.CV↗