Bone Metastases Lesion Segmentation on Breast Cancer Bone Scan Images with Negative Sample Training

The use of deep learning methods for the automatic detection and quantification bone metastases in scan images holds significant clinical value. A fast accurate automated system segmenting metastatic lesions can assist physicians diagnosis. In this study, a small internal dataset comprising 100 breast cancer patients (90 cases metastasis 10 non-metastasis) prostate (50 50 was used model training. Initially, all image labels were binary. We Otsu thresholding method or negative mining to generate non-metastasis mask, thereby transforming into three classes. adopted Double U-Net as baseline made modifications its output activation function. changed function SoftMax accommodate multi-class segmentation. Several enhance performance, including background pre-processing remove information, adding samples improve precision, using transfer leverage shared features between two datasets, which enhances model’s performance. performance investigated via 10-fold cross-validation computed on pixel-level scale. best we achieved had precision 69.96%, sensitivity 63.55%, an F1-score 66.60%. Compared model, represents 8.40% improvement 0.56% sensitivity, 4.33% F1-score. developed has potential provide pre-diagnostic reports final decisions calculation index (BSI) with combination skeleton