TY - JOUR T1 - Evaluation of Deep Learning Based PET Image Enhancement Method in Diagnosis of Lymphoma JF - Journal of Nuclear Medicine JO - J Nucl Med SP - 431 LP - 431 VL - 61 IS - supplement 1 AU - Feng Xu AU - Boyang Pan AU - Xiaochun Zhu AU - Praveen Gulaka AU - Lei Xiang AU - Enhao Gong AU - Tao Zhang AU - Jiazheng Wang AU - Liangjie Lin AU - Yubo Ma AU - Nan-Jie Gong Y1 - 2020/05/01 UR - http://jnm.snmjournals.org/content/61/supplement_1/431.abstract N2 - 431Purpose: This study aims to evaluate the performance of deep learning enhancement (SubtlePET) method on low dose PET images from both quantitative and qualitative perspectives. The impact of deep learning enhancement on clinical decision making was also assessed. Method and Materials: 20 subjects (12 males, mean age: 59 years, age range: 26-79 years, mean body mass index: 22.6, at least 1 FDG positive lesion) underwent clinical whole-body PET/CT exams. PET images were reconstructed using 50%, 75% and 100% counts, respectively. Images reconstructed from 50% and 75% counts were further enhanced using a deep learning (DL) based software (SubtlePET, Subtle Medical, Menlo Park, CA). One nuclear medicine physician reviewed the standard acquisition PET images in Intellispace Portal (Philips, Best, the Netherlands) and identified possible lesions. Volume of interests (VOIs) of the lesions and regions of interest (ROIs) in liver and lung were drawn by the same physician on the standard acquisition. Same lesions, VOIs and ROIs were also reviewed on the DL-enhanced images. Quantitative mean and maximum SUV values per VOI of the standard and DL-enhanced noisy acquisitions were assessed using Bland-Altman test They were further compared using concordance correlation coefficients (CCC) and linear regressions. Qualitative image quality of unlabeled images were evaluated using 5-point Likert scale by two physicians. Evaluation metrics included image quality, diagnostic confidence, and intra-reader repeatability test. The lymphoma staging was evaluated using Lugano staging system and the Deauville scale. Results: A total of 56 lesions were identified with VOIs drawn in the standard acquisition and DL-enhanced PET images. The Bland-Altman plots showed minimal differences of SUVs between standard acquisition and DL-enhanced PET images. In SUVmax, CCC values were 0.94 and 0.93 for 75% counts and 50% counts respectively, as compared to standard of care (SoC) acquisition. Pearson coefficient values were 0.99 and 0.99 for 75% counts and 50% counts respectively, in reference to SoC acquisition. These results indicated very strong agreement between SoC and DL-enhanced scans. In SUVmean, CCC values were both 0.86 for 75% counts and 50% counts respectively. Pearson coefficient values were 0.97 and 0.98 for 75% counts and 50% counts respectively. These results also indicated good agreement between the SoC and DL-enhanced scans. SoC were voted only slightly better than DL-enhanced images in general image quality and detail of FDG distribution while the DL-enhanced scans with 75% counts had the best interobserver agreement. No artifacts and excellent general diagnostic confidence were found in DL-enhanced images, which were the same as SoC. The lymphoma staging of the 20 patients assessed by the two physicians according to Lugano system were the same between DL-enhanced acquisitions and SoC acquisitions. Deauville scales of all 56 lesions were also the same between DL-enhanced acquisitions and SoC acquisitions. Conclusions: Deep learning can help accelerate PET acquisitions by 25% or even 50% without compromising quantitative SUV values and qualitative image quality as compared to the standard duration acquisitions. Clinical relevance/application: Deep learning can enhance image quality of noisy accelerated PET acquisitions thereby enabling higher comfort for patients, higher throughput of PET scans for hospitals, or reduced radiotracer dosages. Figure 1: Bland-Altman plots of SUVmax and SUVmean of 56 lesions. Figure 2: linear regression of of SUVmax and SUVmean of 56 lesions between SubtlePET enhanced and SoC. Figure 3: Maximum intensity projection (first three columns), PET images (fourth column) and fused images (fifth column) of a 79-year-old man were reconstructed and enhanced using 100%(A), 75%(B) and 50%(C) counts with SubtlePET. Images reconstructed from 100% and those from 75%, 50% counts with DL-enhancement were the same on visual assessment. View this table:Qualitative Assessment Results of Image Quality ER -