Current students - PhD DADS

WANG HAIYANG

Cycle: XXXVIII

Advisor: MAINARDI LUCA
Tutor: CAIANI ENRICO GIANLUCA

Major Research topic:
Data analysis for diagnosis of oncological patients

Abstract:

Cutting-edge narrow band images (NBI) in medical diagnosis has been conducted. This innovative technique offers a distinct advantage in identifying surface-level mucosal lesions before they become apparent on traditional radiological scans. Subsequently, these lesions can be excised through endoscopic resection. Numerous researchers are currently investigating the potential of utilizing NBI for diagnosing oropharyngeal conditions. Nonetheless, in practical application, the task of delineating lesion boundaries and differentiating between benign and malignant cases remains challenging, particularly for less experienced medical practitioners. Attaining proficiency in this skill requires a considerable investment of training hours. Hence, the implementation of an automated detection analysis model for NBI holds great promise. While it's acknowledged that deep learning models, especially deep neural networks, can incur high computational and resource costs, demanding substantial computational power and training time, their effectiveness cannot be overlooked. Nonetheless, a notable drawback lies in the complexity of understanding the underlying rationales behind these models' decisions, a crucial aspect for their practical clinical utility. So, it is demanding that a robust, efficient, and good performance classification system with an explanation in image lesions. Based on a public dataset used available at https://zenodo.org/record/6674034#.ZAIYx3bML52 . A model based on radiomics is proposed to distinguish patients with malignant and benign laryngeal lesions. It could be more practical and efficient in clinical practice of future. Firstly, a serial of necessary steps to preprocess the data including data checking and data cleaining. Then radiomics was utilized to extract the features of data. In the classification step, the performance of the following classifiers were investigated: support vector machine (SVM), Random Forest, k-nearest neighbors algorithm （KNN ）and Gradient-Boosting classifiers. The results were analyzed by a series of performance metrics, namely: accuracy, precision, recall, F1 score and area under cure (AUC). While positive result is shown. it is essential to evaluate a broader range of datasets to effectively mitigate any biases. Simultaneously, there remains a pressing need for enhancing the model's interpretability to cater to physicians' understanding. My research endeavors are continuously evolving. Presently, I am engaged in the exploration of a fresh dataset obtained from Brescia Hospital. The objective is to establish a resilient and precise data analysis system that offers transparency in the diagnosis of Narrow Band Imaging (NBI), particularly concerning oncological patients.