Estimating subjective evaluation of low-contrast resolution using convolutional neural networks

Yujiro Doi, Atsushi Teramoto, Ayumi Yamada, Masanao Kobayashi, Kuniaki Saito, Hiroshi Fujita

Research output: Contribution to journalArticlepeer-review

2 Citations (Scopus)


To develop a convolutional neural network-based method for the subjective evaluation of computed tomography (CT) images having low-contrast resolution due to imaging conditions and nonlinear image processing. Four radiological technologists visually evaluated CT images that were reconstructed using three nonlinear noise reduction processes (AIDR 3D, AIDR 3D Enhanced, AiCE) on a CT system manufactured by CANON. The visual evaluation consisted of two items: low contrast detectability (score: 0–9) and texture pattern (score: 1–5). Four AI models with different convolutional and max pooling layers were constructed and trained on pairs of CANON CT images and average visual assessment scores of four radiological technologists. CANON CT images not used for training were used to evaluate prediction performance. In addition, CT images scanned with a SIEMENS CT system were input to each AI model for external validation. The mean absolute error and correlation coefficients were used as evaluation metrics. Our proposed AI model can evaluate low-contrast detectability and texture patterns with high accuracy, which varies with the dose administered and the nonlinear noise reduction process. The proposed AI model is also expected to be suitable for upcoming reconstruction algorithms that will be released in the future.

Original languageEnglish
Pages (from-to)1285-1296
Number of pages12
JournalPhysical and Engineering Sciences in Medicine
Issue number4
Publication statusPublished - 12-2021

All Science Journal Classification (ASJC) codes

  • Biotechnology
  • Radiological and Ultrasound Technology
  • Biophysics
  • Biomedical Engineering
  • Instrumentation
  • Radiology Nuclear Medicine and imaging


Dive into the research topics of 'Estimating subjective evaluation of low-contrast resolution using convolutional neural networks'. Together they form a unique fingerprint.

Cite this