Data Augmentation and Ensemble Model for Benign and Malignant Diagnosis of Parotid Tumors in Deep Learning
Ryotaro Masaki, Koji Kinoshita, Masaharu Isshiki · 2024
The application of deep learning in medical image diagnosis aims to reduce physicians' workload and enhance diagnostic accuracy. However, challenges such as patient privacy protection and the shortage of annotated medical data hinder the performance of deep learning models. Accurate diagnosis of parotid gland tumors is vital as treatment varies significantly between benign (preservation) and malignant (resection) cases, directly affecting the patient's quality of life (QOL). This study introduces two key methods to achieve high-accuracy diagnosis in data-constrained environments. Firstly, we utilize frequency domain data augmentation techniques, enabling the extraction of diverse image features that enhance model training. Secondly, we implement an ensemble model combining Convolutional Neural Networks (CNN) and Transformer models, leveraging their complementary feature extraction strengths. Our experimental results, using MRI T2 images from 151 patients, indicate that our proposed methods achieve an accuracy of 86%, surpassing traditional models. The frequency domain augmentation and ensemble approach notably improve the classification of parotid gland tumors, highlighting its potential for clinical application.