Multi-branch CNNFormer: a novel framework for predicting prostate cancer response to hormonal therapy

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Oncology
• Biomarker Integration

Background:

• Prostate cancer (PC) treatment response prediction is crucial for personalized therapy.
• Current methods face limitations in capturing complex spatial and global features from multi-modality MRI.
• Integrating clinical markers like prostate-specific antigen (PSA) is essential.

Purpose of the Study:

• To develop and validate a hybrid model combining Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) for accurate PC response prediction.
• To overcome the limitations of individual CNNs (long-range relations) and ViTs (localization) in analyzing PC lesions.
• To integrate multi-modality MRI and PSA data for enhanced prediction accuracy.

Main Methods:

• A 3D multi-branch CNN-Transformer (CNNFormer) model was developed, integrating 3D CNN for detailed feature extraction and 3D ViT for global feature learning.
• The model processes volumetric MRI data to capture both localized details and salient global patterns.
• The framework was evaluated on a cohort of 39 patients, stratified by PSA levels.

Main Results:

• The CNNFormer model achieved high accuracy (97.50%) in distinguishing between responders and non-responders to hormonal therapy.
• Exceptional sensitivity (100%) and specificity (95.83%) were observed, highlighting the model's robust performance.
• The results demonstrate the model's effectiveness even with a small patient cohort.

Conclusions:

• The proposed CNNFormer framework shows significant potential for improving personalized prostate cancer treatment planning and monitoring.
• Combining CNN and ViT strengths offers a robust method for predicting PC response to hormonal therapy.
• This approach can enhance clinical decision-making in managing prostate cancer.