A PERCIST 1.0 based Computer Assisted Detection (CAD) system: Performance compared to an expert reader in the evaluation of breast cancer

Objectives To evaluate the performance of a Computer Assisted Detection (CAD) system based on the PERCIST 1.0 criteria as compared to an expert reader in tumor response assessment in breast cancer.

Methods Baseline FDG PET studies acquired on 18 patients in an IRB approved multi-site clinical trial for treatment of breast cancer were used in this analysis. Each study was evaluated independently by both an expert reader and by a PERCIST 1.0 based CAD program (developed in-house). For each study, the expert reader and CAD system independently applied the PERCIST 1.0 criteria by first measuring a 3cm VOI in the liver and calculating the PERCIST 1.0 threshold for disease detectability. Subsequently, tumor targets were identified. Multiple parameters, including PEAK-SUL and SUL-Max, were recorded for all tumor targets. The results of the CAD driven analysis were then compared against those of the expert reader.

Results The PERCIST 1.0 threshold measurements were not statistically different between the expert reader and the auto-detect algorithm used by the CAD system (t-test p-value = 0.98, ICC = 0.91). When PEAK-SUL was used to identify tumor targets, 15 (83%) studies were determined to be eligible for PERCIST 1.0 assessment, with the expert reader and CAD system performing identically (100% concordance). When using SUL-Max to identify tumor targets, the expert reader again found the same 15 (83%) studies as eligible for assessment whereas the CAD system identified 16 (89%) of the original 18 cases to be eligible (the same 15 as the expert reader, plus an additional case), an improvement over the expert reader by 7%.

Conclusions This work demonstrates that PERCIST 1.0 can be implemented in a CAD tool. It also demonstrates that this tool performs as well as an independent expert reader working alone when evaluating FDG-PET studies in cases of breast cancer. Use of such a tool may increase the efficiency of readers, as well as the objectivity and comparability of results amongst multiple readers