Large Multi-institutional Data Research Articles

Extended nnU-Net for Brain Metastasis Detection and Segmentation in Contrast-Enhanced Magnetic Resonance Imaging With a Large Multi-Institutional Data Set

PurposeTo investigate an extended self-adapting nnU-Net framework for detecting and segmenting brain metastases (BM) on MRI. ApproachSix different nnU-Net systems with adaptive data sampling, adaptive Dice loss (ADL) or different patch/batch sizes were trained and tested for detecting and segmenting intraparenchymal BM with a size ≥ 2 mm on 3D post-Gd T1-weighted MRI volumes using 2092 patients from seven institutions (1712, 195 and 185 patients for training, validation and testing, respectively). Gross tumor volumes (GTVs) of BM delineated by physicians for stereotactic radiosurgery (SRS) were collected retrospectively and curated at each institute. Additional centralized data curation was carried out to create GTVs of uncontoured BM by two radiologists to improve accuracy of ground truth. The training dataset was augmented with synthetic BMs of 1025 MRI volumes using a 3D generative pipeline. BM detection was evaluated by lesion-level sensitivity and false-positive (FP) rate. BM segmentation was assessed by lesion-level Dice similarity coefficient (DSC), 95-percentile Hausdorff distance (HD95) and average HD. The performances were assessed across different BM sizes. Additional testing was performed using a second dataset of 206 patients. ResultsOf the six nnU-Net systems, the nnU-Net with ADL achieved the best detection and segmentation performance on the first testing dataset. At an FP rate of 0.65±1.17, overall sensitivity was 0.904 for all sizes of BM, 0.966 for BM ≥ 0.1 cm3 and 0.824 for BM < 0.1 cm3. Mean values of DSC, HD95 and average HD of all detected BMs were 0.758, 1.45 mm and 0.23 mm, respectively. Performances on the second testing dataset achieved sensitivity of 0.907 at an FP rate of 0.57±0.85 for all BM sizes, and average HD of 0.33 mm for all detected BM. ConclusionsOur proposed extension of the self-configuring nnU-Net framework substantially improved small BM detection sensitivity while maintaining a controlled FP rate. Clinical utility of the extended nnU-Net model for assisting early BM detection and SRS planning will be investigated.

178P Survival benefit of splenic hilar lymph nodes (no.10) dissection in B4 type gastric carcinoma: An IPTW propensity score analysis of large multi-institutional data

178P Survival benefit of splenic hilar lymph nodes (no.10) dissection in B4 type gastric carcinoma: An IPTW propensity score analysis of large multi-institutional data

Deep learning-based prediction of rib fracture presence in frontal radiographs of children under two years of age: a proof-of-concept study.

In this proof-of-concept study, we aimed to develop deep-learning-based classifiers to identify rib fractures on frontal chest radiographs in children under 2 years of age. This retrospective study included 1311 frontal chest radiographs (radiographs with rib fractures, n = 653) from 1231 unique patients (median age: 4 m). Patients with more than one radiograph were included only in the training set. A binary classification was performed to identify the presence or absence of rib fractures using transfer learning and Resnet-50 and DenseNet-121 architectures. The area under the receiver operating characteristic curve (AUC-ROC) was reported. Gradient-weighted class activation mapping was used to highlight the region most relevant to the deep learning models' predictions. On the validation set, the ResNet-50 and DenseNet-121 models obtained an AUC-ROC of 0.89 and 0.88, respectively. On the test set, the ResNet-50 model demonstrated an AUC-ROC of 0.84 with a sensitivity of 81% and specificity of 70%. The DenseNet-50 model obtained an AUC of 0.82 with 72% sensitivity and 79% specificity. In this proof-of-concept study, a deep learning-based approach enabled the automatic detection of rib fractures in chest radiographs of young children with performances comparable to pediatric radiologists. Further evaluation of this approach on large multi-institutional data sets is needed to assess the generalizability of our results. In this proof-of-concept study, a deep learning-based approach performed well in identifying chest radiographs with rib fractures. These findings provide further impetus to develop deep learning algorithms for identifying rib fractures in children, especially those with suspected physical abuse or non-accidental trauma.

Artificial Intelligence and Laryngeal Cancer: From Screening to Prognosis: A State of the Art Review.

This state of the art review aims to examine contemporary advances in applications of artificial intelligence (AI) to the screening, detection, management, and prognostication of laryngeal cancer (LC). Four bibliographic databases were searched: PubMed, EMBASE, Cochrane, and IEEE. A structured review of the current literature (up to January 2022) was performed. Search terms related to topics of AI in LC were identified and queried by 2 independent reviewers. Citations of selected studies and review articles were also evaluated to ensure comprehensiveness. AI applications in LC have encompassed a variety of data modalities, including radiomics, genomics, acoustics, clinical data, and videomics, to support screening, diagnosis, therapeutic decision making, and prognosis. However, most studies remain at the proof-of-concept level, as AI algorithms are trained on single-institution databases with limited data sets and a single data modality. AI algorithms in LC will need to be trained on large multi-institutional data sets and integrate multimodal data for optimal performance and clinical utility from screening to prognosis. Out of the data types reviewed, genomics has the most potential to provide generalizable models thanks to available large multi-institutional open access genomic data sets. Voice acoustic data represent an inexpensive and accurate biomarker, which is easy and noninvasive to capture, offering a unique opportunity for screening and monitoring of LA, especially in low-resource settings.

Comparative study of bilateral axillo-breast approach endoscopic and robotic thyroidectomy: propensity score matching analysis of large multi-institutional data.

PurposeThe aim of this study was to compare the large multi-institutional data of surgical outcomes of bilateral axillo-breast approach (BABA) robotic (RT) and endoscopic thyroidectomy (ET) and to evaluate the merits of robotic thyroidectomy.MethodsFrom 2004 to 2015, 1,029 patients underwent BABA ET, and from 2008 to 2015, 2003 patients underwent BABA RT in 3 large volume centers in Korea. Two groups were retrospectively compared in terms of clinicopathologic characteristics, complications, surgical completeness, and long-term outcomes using propensity score matching analysis.ResultsBoth groups had similar demographic characteristics including age, sex, tumor size, pathologic stage, and hospital stay after matching. Each group had similar rate of transient hypoparathyroidism, however, ET showed significantly higher rate of permanent hypoparathyroidism (ET 5.2% vs. RT 2.3%, P = 0.05), and transient vocal cord palsy (ET 14.4% vs. RT 9.1%, P = 0.006). Total operation time was longer in the ET group irrespective of surgical extents, including lobectomy (P = 0.016), total thyroidectomy (P = 0.031), and total thyroidectomy with central lymph node dissection (P = 0.019). The rate of patients with off-Tg under 1.0 ng/mL after 1st ablation was significantly higher in RT than ET group (ET 64.6% vs. RT 92.7%, P < 0.001). In long-term follow-up of cancer patients, 1.4% experienced recurrence after ET (10 cases), while 0.3% cases experienced recurrence after RT (5 cases) (P < 0.001).ConclusionBoth ET and RT can be safe and effective methods to treat thyroid diseases. However, the application of robotic system may help to overcome the limitations of the instruments and surgeon's skills.

The effect of limited (tertiary) Gleason pattern 5 on the new prostate cancer grade groups

The risk of recurrence for prostatic adenocarcinoma after prostatectomy, as detected by prostate-specific antigen or other modalities, is based primarily on Gleason score along with pathologic tumor stage and surgical margin status. Recent large multi-institutional data spanning the last decade have supported modification of risk of recurrence stratification based on grade groups: grade group 1 (3+3=6), grade group 2 (3+4=7), grade group 3 (4+3=7), grade group 4 (4+4=8), and grade group 5 (Gleason scores 9 and 10). Using currently accepted grading definitions of grade patterns and grading rules, this study examines how the introduction of a limited, less than 5%, Gleason pattern 5 component at prostatectomy affects prognosis and fits into the grade group schema and reporting. The aggregate data from 2 independent major academic medical centers comprising 7606 patient records were analyzed with respect to biochemical recurrence-free survival. The presence of a limited (tertiary) Gleason pattern 5 component in the context of Gleason scores 3+4=7 (grade group 2) and 4+3=7 (grade group 3) imparts an intermediate prognosis relative to the next highest grade group. As such, we suggest that an additional comment and designation to the grade groups be provided reflecting the increased risk of recurrence in such cases (such as grade group 2+ or 3+). In contrast, the presence of limited (<5%) Gleason pattern 5 in the context of Gleason score 4+4=8 imparts a poor prognosis equivalent to grade group 5 and therefore should be reported as grade group 5.

WE-G-BRD-09: Prediction of Local Control/Failure by Using Feature Histogram Selection in Follow-Up T2-Weighted MR Image in Spinal Tumors After Stereotactic Body Radiation Therapy

Purpose: In follow-up T2-weighted MR images of spinal tumor patients treated with stereotactic body radiation therapy (SBRT), high intensity features embedded in dark surroundings may suggest a local failure (LF). We investigated image intensity histogram in imaging features to predict LF and local control (LC). Methods: Sixty-seven spinal tumors were treated with SBRT at our institution with scheduled follow-up MR T2-weighted (TR 3200–6600ms; TE 75-132ms) imaging. The LF group included 10 tumors with 8.7 months median follow-up, while the LC group had 11 tumors with 24.1 months median follow-up. The follow-up images were fused to the planning CT. Image intensity histograms of the GTV were calculated. Voxels in greater than 90% (V90), 80% (V80), and peak (Vpeak) of the histogram were grouped into sub-ROIs to determine the best feature histogram. The intensity of each sub-ROI was evaluated using the mean T2-weighted signal ratio (intensity in sub-ROI / intensity in normal vertebrae). An ROC curve in predicting LF for each sub-ROI was calculated to determine the best feature histogram parameter for LF prediction. Results: Mean T2-weighted signal ratio in the LF group was significantly higher than that in the LC group for all sub-ROIs (1.1±0.4 vs. 0.7±0.2, 1.2±0.4 vs. 0.8±0.2, 1.4±0.5more » vs. 0.8±0.2, for V90, V80, and Vpeak, p=0.02, 0.02, and 0.002, respectively). The corresponding areas-under-curve (AUC) of ROC were 0.78, 0.80, and 0.87, p=0.02, 0.03, 0.004, respectively. No correlation was found between T2-weighted signal ratio in Vpeak and follow-up time (Pearson's ρ=0.15). Conclusion: Increased T2-weighted signal can be used to identify local failure while decreased signal indicates local control after spinal SBRT. By choosing the best histogram parameter (here the Vpeak), the AUC of the ROC can be substantially improved, which implies reliable prediction of LC and LF. These results are being further studied and validated with large multi-institutional data.« less

Comparison of Laparoscopic Radical Nephrectomy and Open Radical Nephrectomy for Pathologic Stage T1 and T2 Renal Cell Carcinoma With Clear Cell Histologic Features: A Multi-institutional Study

To assess the oncologic efficacy of laparoscopic radical nephrectomy (LRN) compared with open radical nephrectomy (ORN) in patients with clear cell renal cell carcinoma (RCC). We analyzed the data from 2561 patients who had undergone radical nephrectomy for RCC at 26 institutions in Korea from June 1998 to December 2007. The clinical data of 631 patients with clear cell RCC in the LRN group were compared with the clinical data of 924 patients in the ORN group. The patients with Stage pT3 or greater and those with lymph node or distant metastases were excluded to avoid a selection bias. To evaluate the technical adequacy and oncologic outcome, we compared the perioperative parameters and 5-year overall and disease-free survival rates. The operative time was significantly longer in the LRN group than in the ORN group (219 ± 77 vs 182 ± 62 minutes, P < .001), but the estimated blood loss and complication rate were significantly lower in the LRN group than in the ORN group (P < .001 and P < .001, respectively). On univariate analysis, the LRN group had 5-year overall (93.5% vs 89.8%, P = .120) and recurrence-free (94.0% vs 92.8%, P = .082) survival rates equivalent to those of the ORN group. Even after adjusting for age, sex, T stage, tumor grade, and body mass index in a Cox proportional hazards model, statistically significant differences between the 2 groups were not found for the 5-year overall (hazard ratio 1.523, P = .157) and recurrence-free (hazard ratio 0.917, P = .773) survival rates. Our large multi-institutional data have shown that LRN provides survival outcomes equivalent to those of ORN in patients with Stage pT1-T2 clear cell RCC.

Prostate Cancer Gene 3 (PCA3): Development and Internal Validation of a Novel Biopsy Nomogram

Urinary prostate cancer gene 3 (PCA3) represents a promising novel marker of prostate cancer detection. To test whether urinary PCA3 assay improves prostate cancer (PCa) risk assessment and to construct a decision-making aid in a multi-institutional cohort with pre-prostate biopsy data. PCA3 assay cut-off threshold analyses were followed by logistic regression models which used established predictors to assess PCa-risk at biopsy in a large multi-institutional data set of 809 men at risk of harboring PCa. Regression coefficients were used to construct four sets of nomograms. Predictive accuracy (PA) estimates of biopsy outcome predictions were quantified using the area under the curve of the receiver operator characteristic analysis in models with and without PCA3. Bootstrap resamples were used for internal validation and to reduce overfit bias. The extent of overestimation or underestimation of the observed PCa rate at biopsy was explored graphically using nonparametric loss-calibration plots. Differences in PA were tested using the Mantel-Haenszel test. Finally, nomogram-derived probability cut-offs were tested to assess the ability to identify patients with or without PCa. PCA3 was identified as a statistically independent risk factor of PCa at biopsy. Addition of a PCA3 assay improved bootstrap-corrected multivariate PA of the base model between 2% and 5%. The highest increment in PA resulted from a PCA3 assay cut-off threshold of 17, where a 5% gain in PA (from 0.68 to 0.73, p=0.04) was recorded. Nomogram probability-derived risk cut-off analyses further corroborate the superiority of the PCA3 nomogram over the base model. PCA3 fulfills the criteria for a novel marker capable of increasing PA of multivariate biopsy models. This novel PCA3-based nomogram better identifies men at risk of harboring PCa and assists in deciding whether further evaluation is necessary.

The RACHS-1 risk categories reflect mortality and length of stay in a Danish population of children operated for congenital heart disease☆

The Risk Adjusted classification for Congenital Heart Surgery (RACHS-1) was created in order to compare in-hospital mortality for groups of children undergoing surgery for congenital heart disease. The method was evaluated with two large multi-institutional data sets-the Paediatric Cardiac Care Consortium (PCCC) and Hospital Discharge (HD) data from three states in the USA. The RACHS-1 classification was later applied to a large German paediatric cardiac surgery population in Bad Oeynhausen (BO), where it was found that the RACHS-1 categories were also associated with length of stay. We applied the RACHS-1 classification to the 957 operations performed during January 1996 to December 2002 at Skejby Sygehus, Denmark and we examined the association between the RACHS-1 categories, in-hospital mortality and length of stay in the Intensive Care Unit. The operations were classified according to the six RACHS-1 categories by matching the procedure of each patient with a risk category. The ability of the RACHS-1 classification to predict mortality in our population was examined by estimating the area under the receiver operator characteristic (ROC) curve. Likelihood ratio chi(2) tests were used to compare the distribution of RACHS-1 categories and the distribution of mortality with PCCC, HD and BO. Linear regression was used to examine the correlation between the RACHS-1 categories and length of stay in the Intensive Care Unit. The RACHS-1 category frequencies in our population were: category 1: 18.4%, category 2: 37.4%, category 3: 34.6%, category 4: 8.2%, category 5: 0% and category 6: 1.5%. The overall ability of the RACHS-1 classification to predict in-hospital mortality (area under the ROC curve 0.741; 95% confidence interval=0.690; 0.791) was equal to the findings from larger populations. We found no differences in the category specific mortality when comparing with the larger reported series. There was a positive association between RACHS-1 category and length of stay in the Intensive Care Unit. The RACHS-1 classification can also be used to predict in-hospital mortality and length of stay in the Intensive Care Unit in a small volume centre.

Familiar/interactive raters are not always best: The influence of sampling schedules and class of behavior

Retrospective rating scales are widely used for formal assessment of typical performance. Raters who are the most familiar/interactive with ratees are routinely recommended to maximize the quality of ratings. This caveat to use the most familiar/interactive raters fails to distinguish sampling parameters of the observations on which ratings are based that may be important to assessing different classes of behavior. We hypothesized that systematic observational schedules would be of greater importance to ratings of public events than familiarity/interaction, per se, while the caveat would hold for ratings of private events. We used the Psychotic Inpatient Profile (PIP), which provides separate factor scores for ratings of public and private events, to examine these hypotheses in a quasi-experimental study with adult inpatients of mental hospitals. A large multiinstitutional data set provided retrospective PIP ratings by two types of raters. The most familiar/interactive local clinical staff for each client completed the PIP after observing on an ad lib schedule, along with ongoing job duties. Unfamiliar, noninteractive raters completed the PIP for each client after observing on a systematic time-sampling schedule for purposes of coding an entirely different instrument. Data were selected so that each of 189 clients received PIP scores from four raters, reflecting functioning during the same time period based on day-shift observations by one rater of each type and evening-shift observations by one rater of each type. Analyses of variance, consistency/discriminability of ratings, and prediction of social-action outcomes all supported the hypotheses. We discuss alternative strategies that are better for assessing typical performance in most circumstances. We also provide recommendations for improving the adequacy of observations for those circumstances in which the standardized retrospective rating scale could be a cost-effective assessment strategy.