Tumour segmentation, tracking and analysis in intra-subject whole-body PET-CT scans of lymphoma patients

Malignant lymphoma is a severe threat to public health in most countries. Positron Emission Tomography - Computed Tomography (PET-CT) scanning has replaced PET-only scanning in clinical routine to diagnose diverse types of cancer and it fundamentally supports the clinical evaluation of lymphoma progression and therapy response. In order to automatically process a large scale of PET-CT scans, a prototype system consisting of automated lesion segmentation, computerised lesion tracking and dynamic online reports has been developed. Accurate segmentation of heterogeneous carcinoma lesions in medical images is vital to the treatment planning, assessment of therapy response and other oncological applications. A modified variational Bayes inference (VBI) model with clinical meaningful initialisation is developed to simultaneously delineate lesion boundaries in both anatomical and functional images. The experimental results evaluated by heterogeneous clinical lymphoma lesions show less error rate (6.11% ± 4.92%) and better model capabilities than the alternative method. Accurate assessment of patients' response to therapy is critical for the cancer treatment planning in clinical practices. Tracking the corresponding relationship of lesions between a pair of PET-CT scans of the same patient is the basic step in treatment response assessment. A computerised method is developed to model the lesion changes between successive scans through a corresponding matrix. This matrix's values quantitatively describe the five types of computer-recognised lesion changes between successive scans. Comparing with the doctors' diagnostic reports for the examined clinical PET­CT scans, the proposed approach accurately detects and tracks all of 38-lesions changes recorded by doctors. In order to illustrate the clinical application of the proposed method, this research developed a prototype system embedding the functions of lesion segmentation, lesion feature extraction and feature analysis into a computational procedure presenting dynamic online reports to medical end­users. Once a lesion is identified, the end-users will immediately receive the updated analytical reports of lesion features. In summary, the modified VBI model initialised by functional images is capable of delineating more accurate lesion boundaries in heterogeneous clinical datasets. The lesion tracking strategy provides a computerised method using the lesion segmentation results to assess the lesion response to treatment. The prototype system demonstrates the feasibility of applying this academic research to clinical practice