Automatic lesion detection in thoracic PET-CT images with region probabilities

Objectives To develop an automatic approach to detect lesions from thoracic FDG PET-CT images.

Methods Based on our previous study [1], regions with SUVs higher than a threshold (TH) are identified as lesions. The regions are created using mean-shift clustering to reduce noise at voxel-level, and TH is defined based on the maximum SUV of a subject and mean SUV of mediastinum. The thresholding results, however, require heuristic refinement to reduce false positive detection. Thus we propose a new probabilistic approach for accurate and robust lesion detection. First, by positioning the minimum SUV, TH and maximum SUV of a subject as the 0, 0.5 and 1 probabilities of being lesion, the probability of a region being lesion is computed by linear interpolation of its SUV in 0~0.5 or 0.5~1 range. Then, a graphical model is constructed with region-wise probabilities as the node costs and SUV similarities between neighboring regions as the edge costs. The graphical model is implemented using Conditional Random Field, and the optimization output gives the region-wise labelling of lesion or background.

Results The method was tested on 40 NSCLC PET-CT studies with 64 lesions, and compared to three SUV threshold-based approaches: (1) SUV-2.5, (2) 50% SUVmax, and (3) (0.15*SUVmax+SUVmed) [1]. A correct detection of lesion with acceptable boundary delineation was considered as true positive. Our method exhibits the highest detection performance. In particular, the designed probabilistic approach achieves high precision (95.4%) while maintaining good recall (96.9%). The precision is about 30%, 13% and 8% higher than the compared techniques (1)-(3), and the recall is about 0%, 11% and 4% higher respectively.

Conclusions The developed method demonstrates higher performance for detecting lesions from thoracic PET-CT images compared to the benchmarked approaches.

Research Support ARC grants.