Training certified detectives to track down the intrinsic shortcuts in COVID-19 chest x-ray data sets.

Deep learning faces a significant challenge wherein the trained models often underperform when used with external test data sets. This issue has been attributed to spurious correlations between irrelevant features in the input data and corresponding labels. This study uses the classification of COVID-19 from chest x-ray radiographs as an example to demonstrate that the image contrast and sharpness, which are characteristics of a chest radiograph dependent on data acquisition systems and imaging parameters, can be intrinsic shortcuts that impair the model's generalizability. The study proposes training certified shortcut detective models that meet a set of qualification criteria which can then identify these intrinsic shortcuts in a curated data set.

Correcting statistical CT number biases without access to raw detector counts: Applications to high spatial resolution photon counting CT imaging.

BACKGROUND: Due to the nonlinear nature of the logarithmic operation and the stochastic nature of photon counts (N), sinogram data of photon counting detector CT (PCD-CT) are intrinsically biased, which leads to statistical CT number biases. When raw counts are available, nearly unbiased statistical estimators for projection data were developed recently to address the CT number bias issue. However, for most clinical PCD-CT systems, users' access to raw detector counts is limited. Therefore, it remains a challenge for end users to address the CT number bias issue in clinical applications. PURPOSE: To develop methods to correct statistical biases in PCD-CT without requiring access to raw PCD counts. METHODS: (1) The sample variance of air-only post-log sinograms was used to estimate air-only detector counts, N ¯ 0 $\bar{N}_0$ . (2) If the post-log sinogram data, y, is available, then N of each detector pixel was estimated using N = N ¯ 0 e - y $N = \bar{N}_0 \, \mathrm{e}^{-y}$ . Once N was estimated, a closed-form analytical bias correction was applied to the sinogram. (3) If a patient's post-log sinogram data are not archived, a forward projection of the bias-contaminated CT image was used to perform a first-order bias correction. Both the proposed sinogram domain- and image domain-based bias correction methods were validated using experimental PCD-CT data. RESULTS: Experimental results demonstrated that both sinogram domain- and image domain-based bias correction methods enabled reduced-dose PCD-CT images to match the CT numbers of reference-standard images within [-5, 5] HU. In contrast, uncorrected reduced-dose PCD-CT images demonstrated biases ranging from -25 to 55 HU, depending on the material. No increase in image noise or spatial resolution degradation was observed using the proposed methods. CONCLUSIONS: CT number bias issues can be effectively addressed using the proposed sinogram or image domain method in PCD-CT, allowing PCD-CT acquired at different radiation dose levels to have consistent CT numbers desired for quantitative imaging.

Training certified detectives to track down the intrinsic shortcuts in COVID-19 chest x-ray data sets.

Deep learning faces a significant challenge wherein the trained models often underperform when used with external test data sets. This issue has been attributed to spurious correlations between irrelevant features in the input data and corresponding labels. This study uses the classification of COVID-19 from chest x-ray radiographs as an example to demonstrate that the image contrast and sharpness, which are characteristics of a chest radiograph dependent on data acquisition systems and imaging parameters, can be intrinsic shortcuts that impair the model's generalizability. The study proposes training certified shortcut detective models that meet a set of qualification criteria which can then identify these intrinsic shortcuts in a curated data set.

Diagnosis of Coronavirus Disease 2019 Pneumonia by Using Chest Radiography: Value of Artificial Intelligence.

Background Radiologists are proficient in differentiating between chest radiographs with and without symptoms of pneumonia but have found it more challenging to differentiate coronavirus disease 2019 (COVID-19) pneumonia from non-COVID-19 pneumonia on chest radiographs. Purpose To develop an artificial intelligence algorithm to differentiate COVID-19 pneumonia from other causes of abnormalities at chest radiography. Materials and Methods In this retrospective study, a deep neural network, CV19-Net, was trained, validated, and tested on chest radiographs in patients with and without COVID-19 pneumonia. For the chest radiographs positive for COVID-19, patients with reverse transcription polymerase chain reaction results positive for severe acute respiratory syndrome coronavirus 2 with findings positive for pneumonia between February 1, 2020, and May 30, 2020, were included. For the non-COVID-19 chest radiographs, patients with pneumonia who underwent chest radiography between October 1, 2019, and December 31, 2019, were included. Area under the receiver operating characteristic curve (AUC), sensitivity, and specificity were calculated to characterize diagnostic performance. To benchmark the performance of CV19-Net, a randomly sampled test data set composed of 500 chest radiographs in 500 patients was evaluated by the CV19-Net and three experienced thoracic radiologists. Results A total of 2060 patients (5806 chest radiographs; mean age, 62 years ± 16 [standard deviation]; 1059 men) with COVID-19 pneumonia and 3148 patients (5300 chest radiographs; mean age, 64 years ± 18; 1578 men) with non-COVID-19 pneumonia were included and split into training and validation and test data sets. For the test set, CV19-Net achieved an AUC of 0.92 (95% CI: 0.91, 0.93). This corresponded to a sensitivity of 88% (95% CI: 87, 89) and a specificity of 79% (95% CI: 77, 80) by using a high-sensitivity operating threshold, or a sensitivity of 78% (95% CI: 77, 79) and a specificity of 89% (95% CI: 88, 90) by using a high-specificity operating threshold. For the 500 sampled chest radiographs, CV19-Net achieved an AUC of 0.94 (95% CI: 0.93, 0.96) compared with an AUC of 0.85 (95% CI: 0.81, 0.88) achieved by radiologists. Conclusion CV19-Net was able to differentiate coronavirus disease 2019-related pneumonia from other types of pneumonia, with performance exceeding that of experienced thoracic radiologists. © RSNA, 2021 Online supplemental material is available for this article.