Hierarchical grouped convolution and multi-scale vision transformer for uncertainty-aware early breast cancer detection using mammography

The Hierarchical Grouped Convolution and Multi-Scale Vision Transformer for Uncertainty-Aware Classification (HGCViT-UAC) framework offers a novel approach to early breast cancer detection using mammography. This method addresses the limitations of existing deep learning models by focusing on both local lesion characteristics and global contextual information, while also providing uncertainty quantification and clinically reliable prediction confidence.

The proposed framework begins with several preprocessing steps, including adaptive median filtering, CLAHE enhancement, breast region segmentation, and patient-wise data partitioning. These steps prepare the mammography images for further analysis.

Next, the framework employs hierarchical grouped convolution to extract multi-level features from the images. This technique allows for the capture of both local and global patterns, improving the model's ability to detect breast cancer.

Following this, the model utilizes multi-scale Vision Transformer (ViT) attention to further process the features. The multi-scale attention mechanism enables the model to consider information at different scales, enhancing its ability to capture relevant patterns in the data.

To improve the reliability and clinically applicable nature of the predictions, the framework incorporates adaptive feature fusion and Monte Carlo Dropout-based uncertainty estimation with calibration analysis. This combination of techniques allows the model to provide calibrated uncertainty estimates, which in turn enables clinicians to make more informed decisions about patient care.

The HGCViT-UAC framework was implemented using Python 3.10 and PyTorch 2.0. It was evaluated on the CBIS-DDSM dataset, which contains 10,239 mammography images from 1,566 subjects. The dataset was split into a patient-wise 70/15/15 split, ensuring that the model was tested on a diverse range of cases.

The results demonstrate the effectiveness of the proposed approach. The HGCViT-UAC model achieved a remarkable 98.67% accuracy, surpassing the baseline ViT model, which had an accuracy of 84.32%. This represents an improvement of 14.35% in classification performance. The model also achieved excellent scores on other evaluation metrics, including an F1-score of 98.62%, an AUC of 0.9914, and an Expected Calibration Error of 0.019. Additionally, the Brier Score of 0.026 indicates good calibration of the model's predictions.

Furthermore, the mean accuracy of the HGCViT-UAC model was 98.67% with a 95% confidence interval of 98.06–99.28%. This wide confidence interval demonstrates the model's robustness and consistency across different cases.

In summary, the HGCViT-UAC framework offers a significant advancement in early breast cancer detection using mammography. By combining hierarchical grouped convolution, multi-scale Vision Transformer attention, adaptive feature fusion, and calibrated uncertainty estimation, the model achieves superior performance compared to existing methods.

This approach holds great promise for improving the reliability and clinical applicability of breast cancer diagnosis, potentially leading to earlier detection and better patient outcomes.