Pharmacokinetics-guided breast tumor segmentation method
Ke Lin, Xiangfeng Wen, Zhisong Qin, Wenying Chen · 2025
Rich pathophysiological information of tumors can be provided by dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI), which has become an essential imaging modality for breast tumor diagnosis. The dynamic metabolic process of contrast agents can be quantitatively characterized using pharmacokinetic parameters. To fully exploit the pathological information embedded in DCE-MRI, a pharmacokinetics-guided breast tumor segmentation method based on an improved U-Net neural network is proposed. Multiphase DCE-MRI images and corresponding PK parameter maps are employed as inputs. Spatial features are extracted by residual convolutions in the encoder, and temporal characteristics are modeled using convolutional long short-term memory (ConvLSTM) networks. Additionally, a channel attention mechanism is incorporated into skip connections to enhance critical feature representation. Spatial resolution is progressively restored by the decoder to generate precise breast tumor segmentation masks. The proposed method was evaluated on a publicly available breast DCE-MRI dataset BreastDM, achieving a Dice coefficient of $\mathbf{8 3. 5 \%}$ and an IoU of $\mathbf{7 2. 1 \%}$ on the test set, representing respective improvements of 4.97% and 4.95% over the traditional U-Net baseline. These results demonstrate the effectiveness of integrating pharmacokinetic parameters into deep learning-based tumor segmentation. The related codes are publicly available at PKGnet.git.