Variational Bayes Inference Based Segmentation of Heterogeneous Lymphoma Volumes in Dual-Modality PET-CT Images

Accurate segmentation of heterogeneous carcinoma lesions in medical images is vital to the treatment planning, assessment of therapy response and other oncological applications. With current state-of-the-art imaging modalities, the CT images enhance the interpretation of cancer functional abnormalities. We applied the variational Bayes inference (VBI) model on both anatomical and functional information for delineating lesion boundary. The model is improved by clinical meaningful initialisation. Clinical data consisting of eight lesions with inhomogeneous carcinoma distribution were used to evaluate the model accuracy. Our algorithm is capable of isolating lesions from background with higher accuracy comparing to the wildly used threshold (40% of SUVmax). The VBI segmentation error is less than 6.11% ± 4.92% which is much better than the results performed by fixed threshold method. The experimental results show that our novel statistic method can produce more accurate segmentation of heterogeneous lymphoma volume in PET-CT images.