Breast Tumor Detection and Classification in Mammogram Images Using Modified YOLOv5 Network

Area of Science:

• Medical Imaging
• Computer Vision
• Oncology

Background:

• Breast cancer incidence is rising, making early diagnosis crucial for recovery.
• Mammography is a key screening tool, but accurate tumor detection and classification remain challenging.
• Existing methods often suffer from high false positive/negative ratios and low accuracy.

Purpose of the Study:

• To develop an advanced deep learning model for accurate breast tumor detection and classification.
• To overcome limitations of previous studies by reducing false positive ratio (FPR) and false negative ratio (FNR) while increasing Matthews correlation coefficient (MCC).

Main Methods:

• A modified YOLOv5 network was employed for breast tumor detection and classification.
• Preprocessing involved image enhancement and removal of artifacts from the CBIS-DDSM dataset.
• The model was trained, validated, and tested on augmented data using specific hyperparameters (batch size 8, learning rate 0.01, momentum 0.843, 300 epochs).

Main Results:

• The proposed YOLOv5 model achieved 96% mAP, 96.50% accuracy, 93.50% MCC, 0.04 FPR, and 0.03 FNR.
• Performance was superior when compared to YOLOv3 and Faster RCNN models.
• The model demonstrated significant improvements in reducing FPR and FNR and increasing MCC.

Conclusions:

• The modified YOLOv5 model effectively detects and classifies breast tumors in mammograms.
• This approach addresses previous research limitations, offering enhanced diagnostic performance for breast cancer screening.
• The findings suggest a promising tool for improving the accuracy and reliability of mammographic analysis.