Application of a validated prostate MRI deep learning system to independent same-vendor multi-institutional data: demonstration of transferability.

OBJECTIVES: To evaluate a fully automatic deep learning system to detect and segment clinically significant prostate cancer (csPCa) on same-vendor prostate MRI from two different institutions not contributing to training of the system. MATERIALS AND METHODS: In this retrospective study, a previously bi-institutionally validated deep learning system (UNETM) was applied to bi-parametric prostate MRI data from one external institution (A), a PI-RADS distribution-matched internal cohort (B), and a csPCa stratified subset of single-institution external public challenge data (C). csPCa was defined as ISUP Grade Group ≥ 2 determined from combined targeted and extended systematic MRI/transrectal US-fusion biopsy. Performance of UNETM was evaluated by comparing ROC AUC and specificity at typical PI-RADS sensitivity levels. Lesion-level analysis between UNETM segmentations and radiologist-delineated segmentations was performed using Dice coefficient, free-response operating characteristic (FROC), and weighted alternative (waFROC). The influence of using different diffusion sequences was analyzed in cohort A. RESULTS: In 250/250/140 exams in cohorts A/B/C, differences in ROC AUC were insignificant with 0.80 (95% CI: 0.74-0.85)/0.87 (95% CI: 0.83-0.92)/0.82 (95% CI: 0.75-0.89). At sensitivities of 95% and 90%, UNETM achieved specificity of 30%/50% in A, 44%/71% in B, and 43%/49% in C, respectively. Dice coefficient of UNETM and radiologist-delineated lesions was 0.36 in A and 0.49 in B. The waFROC AUC was 0.67 (95% CI: 0.60-0.83) in A and 0.7 (95% CI: 0.64-0.78) in B. UNETM performed marginally better on readout-segmented than on single-shot echo-planar-imaging. CONCLUSION: For same-vendor examinations, deep learning provided comparable discrimination of csPCa and non-csPCa lesions and examinations between local and two independent external data sets, demonstrating the applicability of the system to institutions not participating in model training. CLINICAL RELEVANCE STATEMENT: A previously bi-institutionally validated fully automatic deep learning system maintained acceptable exam-level diagnostic performance in two independent external data sets, indicating the potential of deploying AI models without retraining or fine-tuning, and corroborating evidence that AI models extract a substantial amount of transferable domain knowledge about MRI-based prostate cancer assessment. KEY POINTS: • A previously bi-institutionally validated fully automatic deep learning system maintained acceptable exam-level diagnostic performance in two independent external data sets. • Lesion detection performance and segmentation congruence was similar on the institutional and an external data set, as measured by the weighted alternative FROC AUC and Dice coefficient. • Although the system generalized to two external institutions without re-training, achieving expected sensitivity and specificity levels using the deep learning system requires probability thresholds to be adjusted, underlining the importance of institution-specific calibration and quality control.

Influence of a Deep Learning Noise Reduction on the CT Values, Image Noise and Characterization of Kidney and Ureter Stones.

Deep-learning (DL) noise reduction techniques in computed tomography (CT) are expected to reduce the image noise while maintaining the clinically relevant information in reduced dose acquisitions. This study aimed to assess the size, attenuation, and objective image quality of reno-ureteric stones denoised using DL-software in comparison to traditionally reconstructed low-dose abdominal CT-images and evaluated its clinical impact. In this institutional review-board-approved retrospective study, 45 patients with renal and/or ureteral stones were included. All patients had undergone abdominal CT between August 2019 and October 2019. CT-images were reconstructed using the following three methods: filtered back-projection, iterative reconstruction, and PixelShine (DL-software) with both sharp and soft kernels. Stone size, CT attenuation, and objective image quality (signal-to-noise ratio (SNR), contrast-to-noise ratio (CNR)) were evaluated and compared using Bonferroni-corrected Friedman tests. Objective image quality was measured in six regions-of-interest. Stone size ranged between 4.4 × 3.1−4.4 × 3.2 mm (sharp kernel) and 5.1 × 3.8−5.6 × 4.2 mm (soft kernel). Mean attenuation ranged between 704−717 Hounsfield Units (HU) (soft kernel) and 915−1047 HU (sharp kernel). Differences in measured stone sizes were ≤1.3 mm. DL-processed images resulted in significantly higher CNR and SNR values (p < 0.001) by decreasing image noise significantly (p < 0.001). DL-software significantly improved objective image quality while maintaining both correct stone size and CT-attenuation values. Therefore, the clinical impact of stone assessment in denoised image data sets remains unchanged. Through the relevant noise suppression, the software additionally offers the potential to further reduce radiation exposure.

Influence of a novel deep-learning based reconstruction software on the objective and subjective image quality in low-dose abdominal computed tomography.

OBJECTIVES: Modern reconstruction and post-processing software aims at reducing image noise in CT images, potentially allowing for a reduction of the employed radiation exposure. This study aimed at assessing the influence of a novel deep-learning based software on the subjective and objective image quality compared to two traditional methods [filtered back-projection (FBP), iterative reconstruction (IR)]. METHODS: In this institutional review board-approved retrospective study, abdominal low-dose CT images of 27 patients (mean age 38 ± 12 years, volumetric CT dose index 2.9 ± 1.8 mGy) were reconstructed with IR, FBP and, furthermore, post-processed using a novel software. For the three reconstructions, qualitative and quantitative image quality was evaluated by means of CT numbers, noise, signal-to-noise ratio (SNR), and contrast-to-noise ratio (CNR) in six different ROIs. Additionally, the reconstructions were compared using SNR, peak SNR, root mean square error and mean absolute error to assess structural differences. RESULTS: On average, CT numbers varied within 1 Hounsfield unit (HU) for the three assessed methods in the assessed ROIs. In soft tissue, image noise was up to 42% lower compared to FBP and up to 27% lower to IR when applying the novel software. Consequently, SNR and CNR were highest with the novel software. For both IR and the novel software, subjective image quality was equal but higher than the image quality of FBP-images. CONCLUSION: The assessed software reduces image noise while maintaining image information, even in comparison to IR, allowing for a potential dose reduction of approximately 20% in abdominal CT imaging. ADVANCES IN KNOWLEDGE: The assessed software reduces image noise by up to 27% compared to IR and 48% compared to FBP while maintaining the image information.The reduced image noise allows for a potential dose reduction of approximately 20% in abdominal imaging.

Joint Imaging Platform for Federated Clinical Data Analytics.

PURPOSE: Image analysis is one of the most promising applications of artificial intelligence (AI) in health care, potentially improving prediction, diagnosis, and treatment of diseases. Although scientific advances in this area critically depend on the accessibility of large-volume and high-quality data, sharing data between institutions faces various ethical and legal constraints as well as organizational and technical obstacles. METHODS: The Joint Imaging Platform (JIP) of the German Cancer Consortium (DKTK) addresses these issues by providing federated data analysis technology in a secure and compliant way. Using the JIP, medical image data remain in the originator institutions, but analysis and AI algorithms are shared and jointly used. Common standards and interfaces to local systems ensure permanent data sovereignty of participating institutions. RESULTS: The JIP is established in the radiology and nuclear medicine departments of 10 university hospitals in Germany (DKTK partner sites). In multiple complementary use cases, we show that the platform fulfills all relevant requirements to serve as a foundation for multicenter medical imaging trials and research on large cohorts, including the harmonization and integration of data, interactive analysis, automatic analysis, federated machine learning, and extensibility and maintenance processes, which are elementary for the sustainability of such a platform. CONCLUSION: The results demonstrate the feasibility of using the JIP as a federated data analytics platform in heterogeneous clinical information technology and software landscapes, solving an important bottleneck for the application of AI to large-scale clinical imaging data.

Development of size-specific institutional diagnostic reference levels for computed tomography protocols in neck imaging.

PURPOSE: To develop size-specific institutional diagnostic reference levels (DRLs) for computed tomography (CT) protocols used in neck CT imaging (cervical spine CT, cervical CT angiography (CTA) and cervical staging CT) and to compare institutional to national DRLs. MATERIALS AND METHODS: Cervical CT examinations (spine, n = 609; CTA, n = 505 and staging CT, n = 184) performed between 01/2016 and 06/2017 were included in this retrospective study. For each region and examination, the volumetric CT dose index (CTDIvol) and dose-length product (DLP) were determined and binned into size bins according to patient water-equivalent diameter (dw). Linear regression analysis was performed to calculate size-specific institutional DRLs for CTDIvol and DLP, applying the 75th percentile as the upper limit for institutional DRLs. The mean institutional CTDIvol and DLP were compared to national DRLs (CTDIvol 20 mGy for cervical spine CT (DLP 300 mGycm) and cervical CTA (DLP 600 mGycm), and CTDIvol 15 mGy for cervical staging CT (DLP 330 mGycm)). RESULTS: The mean CTDIvol and DLP (±standard deviation) were 15.2 ± 4.1 mGy and 181.5 ± 88.3 mGycm for cervical spine CT; 8.1 ± 4.3 mGy and 280.2 ± 164.3 mGycm for cervical CTA; 8.6 ± 1.9 mGy and 162.8 ± 85.0 mGycm for cervical staging CT. For all CT protocols, there was a linear increase in CTDIvol and DLP with increasing dw. For the CTDIvol, size-specific institutional DRLs increased with dw from 14 to 29 mGy for cervical spine CT, from 5 to 17 mGy for cervical CTA and from 8 to 13 mGy for cervical staging CT. For the DLP, size-specific institutional DRLs increased with dw from 130 to 510 mGycm for cervical spine CT, from 140 to 640 mGycm for cervical CTA and from 140 to 320 mGycm for cervical staging CT. Institutional DRLs were lower than national DRLs by 81% and 67% for cervical spine CT (dw = 17.8 cm), 43% and 51% for cervical CTA (dw = 19.5 cm) and 59% and 53% for cervical staging CT (dw = 18.8 cm) for CTDIvol and DLP, respectively. CONCLUSION: Size-specific institutional DRLs were generated for neck CT examinations. The mean institutional CTDIvol and DLP values were well below national DRLs.

Implementation of Institutional Size-Specific Diagnostic Reference Levels for CT Angiography.

RATIONALE AND OBJECTIVES: To generate institutional size-specific diagnostic reference levels (DRLs) for computed tomography angiography (CTA) examinations and assess the potential for dose optimization compared to size-independent DRLs. MATERIALS AND METHODS: CTA examinations of the aorta, the pulmonary arteries and of the pelvis/lower extremity performed between January 2016 and January 2017 were included in our retrospective study. Water equivalent diameter (Dw) was automatically calculated for each patient. The relationship between Dw and computed tomography dose index (CTDIvol) was analyzed and the 75th percentile was chosen as the upper limit for institutional DRLs. Size-specific institutional DRLs were compared to national size-independent DRLs from Germany and the UK. RESULTS: A total of 1344 examinations were included in our study (n = 733 aortic CTA, n = 406 pulmonary CTA, n = 205 pelvic/lower extremity CTA). Mean Dw was 26 ± 9 cm and mean CTDIvol was 7.0 ± 4.6 mGy. For all CTA protocols, there was a linear progression of CTDIvol with increasing Dw with an R²â€¯= 0.95 in aortic CTA, R²â€¯= 0.94 in pulmonary CTA and R²â€¯= 0.93 in pelvic/lower extremity CTA. Median CTDIvol increased by 0.57 mGy per additional cm Dw in aortic CTA, by 1.1 mGy in pulmonary CTA and by 0.31 mGy in pelvic/lower extremity CTA. Institutional DRLs were lower than national DRLs for average size patients (aortic CTA: Dw 28.2 cm, CTDIvol 7.6 mGy; pulmonary CTA, Dw 27.9 cm, CTDIvol 11.8 mGy; pelvic/lower extremity CTA, Dw 20.0 cm, CTDIvol 6.4 mGy). More dose outliers in small patients were detected with size-specific DRLs compared to national size-independent DRLs (56.4% vs 16.2%). CONCLUSION: We implemented institutional size-specific DRLs for CTA examinations which enabled a more precise analysis compared to national sizeindependent DRLs.

Potential of a machine-learning model for dose optimization in CT quality assurance.

OBJECTIVES: To evaluate machine learning (ML) to detect chest CT examinations with dose optimization potential for quality assurance in a retrospective, cross-sectional study. METHODS: Three thousand one hundred ninety-nine CT chest examinations were used for training and testing of the feed-forward, single hidden layer neural network (January 2016-December 2017, 60% male, 62 ± 15 years, 80/20 split). The model was optimized and trained to predict the volumetric computed tomography dose index (CTDIvol) based on scan patient metrics (scanner, study description, protocol, patient age, sex, and water-equivalent diameter (DW)). The root mean-squared error (RMSE) was calculated as performance measurement. One hundred separate, consecutive chest CTs were used for validation (January 2018, 60% male, 63 ± 16 years), independently reviewed by two blinded radiologists with regard to dose optimization, and used to define an optimal cutoff for the model. RESULTS: RMSE was 1.71, 1.45, and 1.52 for the training, test, and validation dataset, respectively. The scanner and DW were the most important features. The radiologists found dose optimization potential in 7/100 of the validation cases. A percentage deviation of 18.3% between predicted and actual CTDIvol was found to be the optimal cutoff: 8/100 cases were flagged as suboptimal by the model (range 18.3-53.2%). All of the cases found by the radiologists were identified. One examination was flagged only by the model. CONCLUSIONS: ML can comprehensively detect CT examinations with dose optimization potential. It may be a helpful tool to simplify CT quality assurance. CT scanner and DW were most important. Final human review remains necessary. A threshold of 18.3% between the predicted and actual CTDIvol seems adequate for CT quality assurance. KEY POINTS: • Machine learning can be integrated into CT quality assurance to improve retrospective analysis of CT dose data. • Machine learning may help to comprehensively detect dose optimization potential in chest CT, but an individual review of the results by an experienced radiologist or radiation physicist is required to exclude false-positive findings.

Performance and clinical impact of machine learning based lung nodule detection using vessel suppression in melanoma patients.

PURPOSE: To evaluate performance and the clinical impact of a novel machine learning based vessel-suppressing computer-aided detection (CAD) software in chest computed tomography (CT) of patients with malignant melanoma. MATERIALS AND METHODS: We retrospectively included consecutive malignant melanoma patients with a chest CT between 01/2015 and 01/2016. Machine learning based CAD software was used to reconstruct additional vessel-suppressed axial images. Three radiologists independently reviewed a maximum of 15 lung nodules per patient. Vessel-suppressed reconstructions were reviewed independently and results were compared. Follow-up CT examinations and clinical follow-up were used to assess the outcome. Impact of additional nodules on clinical management was assessed. RESULTS: In 46 patients, vessel-suppressed axial images led to the detection of additional nodules in 25/46 (54.3%) patients. CT or clinical follow up was available in 25/25 (100%) patients with additionally detected nodules. 2/25 (8%) of these patients developed new pulmonary metastases. None of the additionally detected nodules were found to be metastases. None of the lung nodules detected by the radiologists was missed by the CAD software. The mean diameter of the 92 additional nodules was 1.5 ±â€¯0.8 mm. The additional nodules did not affect therapeutic management. However, in 14/46 (30.4%) of patients the additional nodules might have had an impact on the radiological follow-up recommendations. CONCLUSION: Machine learning based vessel suppression led to the detection of significantly more lung nodules in melanoma patients. Radiological follow-up recommendations were altered in 30% of the patients. However, all lung nodules turned out to be non-malignant on follow-up.

Tailoring CT Dose to Patient Size: Implementation of the Updated 2017 ACR Size-specific Diagnostic Reference Levels.

RATIONALE AND OBJECTIVES: To use an automatic computed tomography (CT) dose monitoring system to analyze the institutional chest and abdominopelvic CT dose data as regards the updated 2017 American College of Radiology (ACR) diagnostic reference levels (DRLs) based on water-equivalent diameter (Dw) and size-specific dose estimates (SSDE) to detect patient-size subgroups in which CT dose can be optimized. MATERIALS AND METHODS: All chest CT examinations performed between July 2016 and April 2017 with and without contrast material, CT of the pulmonary arteries, and abdominopelvic CT with and without contrast material were included in this retrospective study. Dw and SSDE were automatically calculated for all scans using a previously validated in-house developed Matlab software and stored into our CT dose monitoring system. CT dose data were analyzed as regards the updated ACR DRLs (size groups: 21-25 cm, 25-29 cm, 29-33 cm, 33-37 cm, 37-41 cm). SSDE and volumetric computed tomography dose index (CTDIvol) were used as CT dose parameter. RESULTS: Overall, 30,002 CT examinations were performed in the study period, 3860 of which were included in the analysis (mean age 62.1 ± 16.4 years, Dw 29.0 ± 3.3 cm; n = 577 chest CT without contrast material, n = 628 chest CT with contrast material, n = 346 CT of chest pulmonary, n = 563 abdominopelvic CT without contrast material, n = 1746 abdominopelvic CT with contrast material). Mean SSDE and CTDIvol relative to the updated DRLs were 43.3 ± 26.4 and 45.1 ± 27.9% for noncontrast chest CT, 52.3 ± 23.1 and 52.0 ± 23.1% for contrast-enhanced chest CT, 68.8 ± 29.5 and 70.0 ± 31.0% for CT of pulmonary arteries, 41.9 ± 29.2 and 43.3 ± 31.3% for noncontrast abdominopelvic CT, and 56.8 ± 22.2 and 58.8 ± 24.4% for contrast-enhanced abdominopelvic CT. Lowest dose compared to the DRLs was found for the Dw group of 21-25 cm in noncontrast abdominopelvic CT (SSDE 30.4 ± 21.8%, CTDIvol 30.8 ± 21.4%). Solely the group of patients with a Dw of 37-41 cm undergoing noncontrast abdominopelvic CT exceeded the ACR DRL (SSDE 100.3 ± 59.0%, CTDIvol 107.1 ± 63.5%). CONCLUSIONS: On average, mean SSDE and CTDIvol of our institutional chest and abdominopelvic CT protocols were lower than the updated 2017 ACR DRLs. Size-specific subgroup analysis revealed a wide variability of SSDE and CTDIvol across CT protocols and patient size groups with a transgression of DRLs in noncontrast abdominopelvic CT of large patients (Dw 37-41 cm).

Institutional computed tomography diagnostic reference levels based on water-equivalent diameter and size-specific dose estimates.

Size-specific institutional diagnostic reference levels (DRLs) were generated for chest and abdominopelvic computed tomography (CT) based on size-specific dose estimates (SSDEs) and depending on patients' water-equivalent diameter (Dw). 1690 CT examinations were included in the IRB-approved retrospective study. SSDEs based on the mean water-equivalent diameter of the entire scan volume were calculated automatically. SSDEs were analyzed for different patient sizes and institutional DRLs (iDRLS; 75% percentiles) based on Dw and SSDEs were generated. iDRLs were compared to the national DRLs. Mean volumetric computed tomography dose index (CTDIvol), Dw and SSDEs for all 1690 CT examinations were 7.2 ± 4.0 mGy (0.84-47.9 mGy), 29.0 ± 3.4 cm and 8.5 ± 3.8 mGy (1.2-37.7 mGy), respectively. Overall, the mean SSDEs of all CT examinations were higher than the CTDIvol in chest CT, abdominopelvic CT and upper abdominal CT, respectively (p < 0.001 for all). There was a strong linear correlation between Dw and SSDEs in chest (R2 = 0.66), abdominopelvic (R2 = 0.98) and upper abdominal CT (R2 = 0.96) allowing for the implementation of size-specific institutional DRLs based on SSDEs and patients' Dw. We generated size-specific, Dw-dependent institutional DRLs based on SSDEs, which allow for easier and more comprehensive analyses of CT radiation exposure. Our results indicate that implementation of SSDEs into national DRLs may be beneficial.

ACCURACY OF SIZE-SPECIFIC DOSE ESTIMATE CALCULATION FROM CENTER SLICE IN COMPUTED TOMOGRAPHY.

To evaluate the accuracy of size-specific dose estimate (SSDE) calculation from center slice with water-equivalent diameter (Dw) and effective diameter (Deff). A total of 1812 CT exams (1583 adult and 229 pediatric) were included in this retrospective study. Dw and Deff were automatically calculated for all slices of each scan. SSDEs were calculated with two methods: (1) from the center slice; and (2) from all slices of the volume, which was regarded as the reference standard. Impact of patient weight, height and body mass index (BMI) on SSDE accuracy was assessed. The mean difference between overall SSDE and the center slice approach ranged from 2.0 ± 1.7% (range: 0-15.5%) for pediatric chest to 5.0 ± 3.2% (0-17.2%) for adult chest CT. Accuracy of the center slice SSDE approach correlated with patient size (BMI: r = 0.15-0.43; weight r = 0.26-0.49) which led to SSDE overestimation in small and underestimation in large patients. SSDE calculation using the center slice leads to an error of 2-5%; however, SSDE is underestimated in large patients and overestimation in small patients.