1651. Volumetric Accuracy of Two Deep Learning Based Reconstruction Software for Solid and Sub-solid Nodules in Extremely Low Dose Chest CTs
Authors * Denotes Presenting Author
  1. Thomas Kwack *; Korea University Ansan Hospital
  2. Wooil Kim; Asan Medical Center
  3. Cherry Kim; Korea University Ansan Hospital
  4. Jaehyung Cha; Korea University Ansan Hospital
  5. Zepa Yang; Korea University Guro Hospital
To investigate the accuracy of two deep-learning based reconstruction software for volume measurement of subsolid and solid nodules (SN) at various radiation dosages including those of extremely low dose chest CTs.

Materials and Methods:
Eleven different diameters of SN, ground-glass nodule (GGN), and part-solid nodule (PSN) placed in a phantom were CT scanned at five different radiation dose levels; 120 kVp/435mA [effective dose, 4.08 mSv], 120 kVp/220 mA [2.07 mSv], 120 kVp/90 mA [0.85 mSv], 120 kVp/40 mA [0.38 mSv], and 80 kVp/40 mA [0.12 mSv]. Each CT scan was reconstructed using filtered back projection (FBP), hybrid iterative reconstruction (HIR), and two different deep-learning based reconstruction software (AI1 [vendor-specific] and AI2 [vendor-neutral]). Each nodule volume including GGN, solid and ground-glass opacity (GGO) portions of PSN, and SN was measured semi-automatically, after which absolute percentage measurement errors (APEs) of the measured volumes were calculated. Image noise was calculated to assess the image quality.

Across all nodules and dose settings, the APEs were significantly lower in AI1 and AI2 than in FBP and HIR (all P<0.05). In all radiation dose settings including extremely low dosages, which were equivalent to the posteroanterior and lateral view of the chest radiographs (0.12 mSv), the APEs of GGNs, SN, and inner solid/outer GGO of PSN were lowest in A1 than in A2, FBP, and HIR (all P<0.05). Especially, in 5 mm sized solid portion of PSN and 3 mm sized SN, AI1 show significantly lower APE compared with AI2, FBP, and HIR at all dose settings (all P<0.05). Image noise was significantly lower in AI1 and AI2 than in FBP and HIR for all radiation dose settings (all P<0.05).

Deep-learning based reconstruction software provided the most accurate measurement for volumetry of both subsolid nodules and SN in comparison with FBP and HIR at all radiation dose settings, even in the extremely low dose CT with equivalent to the posteroanterior and lateral view of the chest radiograph. Especially, deep-learning based reconstruction software showed significantly accurate in the small sized SN and solid portion of PSN even in the low dose setting.