Comparison of the Accuracy of a Deep Learning Method for Lesion Detection in PET/CT and PET/MRI Images
Lifang Pang,Zheng Zhang,Guobing Liu,Pengcheng Hu,Shuguang Chen,Yushen Gu,Yukun Huang,Jia Zhang,Yuhang Shi,Tuoyu Cao,Yiqiu Zhang,Hongcheng Shi
DOI: https://doi.org/10.1007/s11307-024-01943-9
2024-08-14
Abstract:Purpose: Develop a universal lesion recognition algorithm for PET/CT and PET/MRI, validate it, and explore factors affecting performance. Procedures: The 2022 AutoPet Challenge's 1014 PET/CT dataset was used to train the lesion detection model based on 2D and 3D fractional-residual (F-Res) models. To extend this to PET/MRI, a network for converting MR images to synthetic CT (sCT) was developed, using 41 sets of whole-body MR and corresponding CT data. 38 patients' PET/CT and PET/MRI data were used to verify the universal lesion recognition algorithm. Image quality was assessed using signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR). Total lesion glycolysis (TLG), metabolic tumor volume (MTV), and lesion count were calculated from the resultant lesion masks. Experienced physicians reviewed and corrected the model's outputs, establishing the ground truth. The performance of the lesion detection deep-learning model on different PET images was assessed by detection accuracy, precision, recall, and dice coefficients. Data with a detection accuracy score (DAS) less than 1 was used for analysis of outliers. Results: Compared to PET/CT, PET/MRI scans had a significantly longer delay time (135 ± 45 min vs 61 ± 12 min) and lower SNR (6.17 ± 1.11 vs 9.27 ± 2.77). However, CNR values were similar (7.37 ± 5.40 vs 5.86 ± 6.69). PET/MRI detected more lesions (with a mean difference of -3.184). TLG and MTV showed no significant differences between PET/CT and PET/MRI (TLG: 119.18 ± 203.15 vs 123.57 ± 151.58, p = 0.41; MTV: 36.58 ± 57.00 vs 39.16 ± 48.34, p = 0.33). A total of 12 PET/CT and 14 PET/MRI datasets were included in the analysis of outliers. Outlier analysis revealed PET/CT anomalies in intestines, ureters, and muscles, while PET/MRI anomalies were in intestines, testicles, and low tracer uptake regions, with false positives in ureters (PET/CT) and intestines/testicles (PET/MRI). Conclusion: The deep learning lesion detection model performs well with both PET/CT and PET/MRI. SNR, CNR and reconstruction parameters minimally impact recognition accuracy, but delay time post-injection is significant.