FS-YOLOv9: A Frequency and Spatial Feature-Based YOLOv9 for Real-time Breast Cancer Detection.

RATIONALE AND OBJECTIVES

Breast cancer screening is critical for reducing mortality rates. YOLOv9, a new real-time object-detection model, is ideal for cancer screening. A customized YOLOv9 model with enhancements for detecting breast cancer on the basis of species and morphological diversity has potential clinical significance.

MATERIALS AND METHODS

The internal dataset consisted of 687 cases split 3:1 for cross-validation. Additionally, 98 cases from external datasets were used for testing. We developed an FS-YOLOv9 model customized for breast cancer detection that incorporated an extra max-pooling layer before the Conv1 of the Adown to enhance high-brightness features. The Adown of the P3 in the backbone was replaced with a high-frequency Haar wavelet convolution kernel, which ignored the low-frequency components during down-sampling to enhance morphology and texture features. The reliability and robustness of our model was determined by measuring the F1 score, the area under curve of free-response receiver operating characteristic (FAUC), mean average precision (mAP), recall, and precision, and comparing them with the findings for the official YOLOv9, YOLOv8, YOLOv5 models.

RESULTS

In comparison with the official YOLOv9 model, FS-YOLOv9 showed a higher average F1 score (0.700 vs. 0.669), FAUC (0.695 vs. 0.662), and mAP50 (0.713 vs. 0.679) in the internal dataset; in the external testing dataset, the FS-YOLOv9 improved the average F1 score, FAUC, and mAP50 by 4.58%, 5.78%, and 4.41% respectively.

CONCLUSION

Our FS-YOLOv9 model showed significantly improved performance in detecting breast cancer, making it more practical for high-risk breast cancer diagnosis.