Liver and Tumor Segmentation in Selective Internal Radiation Therapy 99mTc-MAA SPECT/CT Images using MANet and Histogram Adjustment

Abstract

Selective Internal Radiation Therapy (SIRT) is a widely used radioembolization method for treating primary liver cancer and malignant neoplasms in the liver. Tumor-Liver ratio (TLR) is an important dosimetric parameter for SIRT treatment using 90Y-microspheres. TLR can be calculated from liver and tumor segmentation attained from 99mTc-MAA SPECT/CT. In this study, we propose Multi-Scale Attention U-Net (MANet) and histogram adjustment for accurate liver and tumor segmentation of CT and fused SPECT/CT images, respectively. MANet introduces the multi-scale strategy network to learn and fuse various semantic features from different scales. Histogram adjustment is used for handle normal and abnormal histogram distribution. Noisy-Student pre-trained weights which is learned from noisy images by data augmentation is used in our work. This pre-trained model helps generalize our model and improve overall segmentation performance. 3DIRCADb-01 public dataset is used along with our MAA CT images collected from King Chulalongkorn Memorial Hospital (KCMH) for liver segmentation, and MAA SPECT/CT dataset is used for tumor segmentation. Our proposed method can accurately segment liver, and tumor with DSC of 0.87, 0.65 and IoU of 0.82 and 0.54 respectively.