Hybrid Vision Transformer Model for Accurate Prostate Cancer Classification in MRI Images

Prostate cancer is a widespread disease in men worldwide, posing challenges in early detection due to its diverse nature and the limitations of current diagnostic techniques. In this study, we have developed a new method that combines CNNs with Vision Transformers (ViTs) to improve the classification of prostate cancer using MRI images. To address data imbalance and enhance model robustness, we used dual synthetic oversampling techniques and data augmentation methods. Our preprocessing pipeline reduced noise, preserved edges, and improved local contrast without amplifying noise. For feature extraction, we utilized the GrayLevel Co-occurrence Matrix (GLCM) and shape-based features to represent complex MRI data better. Among the evaluated architectures, ConvNeXt showed the best performance. Using SMOTE, it achieved an F1-score of 97.21% and a MCC of 95.32%. With ADASYN, performance improved further, reaching an F1-score of 98.82% and an MCC of 97.86%. Additionally, we created a web-based application for realtime analysis of prostate MRI scans, which can aid in clinical decision-making. These results demonstrate the potential of our approach to improve the accuracy of prostate cancer diagnosis and provide an interpretable model for clinical use.