This retrospective study aimed to compare the intra- and inter-observer manual-segmentation variability in the feature reproducibility between two-dimensional (2D) and three-dimensional (3D) magnetic-resonance imaging (MRI)-based radiomic features. The study included patients with lipomatous soft-tissue tumors that were diagnosed with histopathology and underwent MRI scans. Tumor segmentation based on the 2D and 3D MRI images was performed by two observers to assess the intra- and inter-observer variability. In both the 2D and the 3D segmentations, the radiomic features were extracted from the normalized images. Regarding the stability of the features, the intraclass correlation coefficient (ICC) was used to evaluate the intra- and inter-observer segmentation variability. Features with ICC > 0.75 were considered reproducible. The degree of feature robustness was classified as low, moderate, or high. Additionally, we compared the efficacy of 2D and 3D contour-focused segmentation in terms of the effects of the stable feature rate, sensitivity, specificity, and diagnostic accuracy of machine learning on the reproducible features. In total, 93 and 107 features were extracted from the 2D and 3D images, respectively. Only 35 features from the 2D images and 63 features from the 3D images were reproducible. The stable feature rate for the 3D segmentation was more significant than for the 2D segmentation (58.9% vs. 37.6%, p = 0.002). The majority of the features for the 3D segmentation had moderate-to-high robustness, while 40.9% of the features for the 2D segmentation had low robustness. The diagnostic accuracy of the machine-learning model for the 2D segmentation was close to that for the 3D segmentation (88% vs. 90%). In both the 2D and the 3D segmentation, the specificity values were equal to 100%. However, the sensitivity for the 2D segmentation was lower than for the 3D segmentation (75% vs. 83%). For the 2D + 3D radiomic features, the model achieved a diagnostic accuracy of 87% (sensitivity, 100%, and specificity, 80%). Both 2D and 3D MRI-based radiomic features of lipomatous soft-tissue tumors are reproducible. With a higher stable feature rate, 3D contour-focused segmentation should be selected for the feature-extraction process.

Objectives The objectives of this study were to develop and test an automated segmentation of R2* iron-overloaded liver images using fuzzy c-mean (FCM) clustering and to evaluate the observer variations. Materials and Methods Liver R2* images and liver iron concentration (LIC) maps of 660 thalassemia examinations were randomly separated into training (70%) and testing (30%) cohorts for development and evaluation purposes, respectively. Two-dimensional FCM used R2* images, and the LIC map was implemented to segment vessels from the parenchyma. Two automated FCM variables were investigated using new echo time and membership threshold selection criteria based on the FCM centroid distance and LIC levels, respectively. The new method was developed on a training cohort and compared with manual segmentation for segmentation accuracy and to a previous semiautomated method, and a semiautomated scheme was suggested to improve unsuccessful results. The automated variables found from the training cohort were assessed for their effectiveness in the testing cohort, both quantitatively and qualitatively (the latter by 2 abdominal radiologists using a grading method, with evaluations of observer variations). A segmentation error of less than 30% was considered to be a successful result in both cohorts, whereas, in the testing cohort, a good grade obtained from satisfactory automated results was considered a success. Results The centroid distance method has a segmentation accuracy comparable with the previous-best, semiautomated method. About 94% and 90% of the examinations in the training and testing cohorts were automatically segmented out successfully, respectively. The failed examinations were successfully segmented out with thresholding adjustment (3% and 8%) or by using alternative results from the previous 1-dimensional FCM method (3% and 2%) in the training and testing cohorts, respectively. There were no failed segmentation examinations in either cohort. The intraobserver and interobserver variabilities were found to be in substantial agreement. Conclusions Our new method provided a robust automated segmentation outcome with a high ease of use for routine clinical application.

In thalassemia patients, R2* liver iron concentration (LIC) measurement is a common clinical tool for assessing iron overload and for determining necessary chelator dose and evaluating its efficacy. Despite the importance of accurate LIC measurement, existing methods suffer from LIC variability, especially at the severe iron overload range due to inclusion of vessel parts in LIC calculation. In this study, we build upon previous Fuzzy C-Mean (FCM) clustering work to formulate a scheme with superior performance in segmenting vessel pixels from the parenchyma. Our method (MIX-FCM) combines our novel 2D-FCM with the existing 1D-FCM algorithm. This study further assessed possible optimal clustering parameters (OP scheme) and proposed a semi-automatic (SA) scheme for routine clinical application. Segmentation of liver parenchyma and vessels was performed on T2* images and their LIC maps in 196 studies from 147 thalassemia major patients. We used manual segmentation as the reference. 1D-FCM clustering was performed on the acquired image alone and 2D-FCM used both the acquired image and its LIC data. To execute the MIX-FCM method, the best outcome (OP-MIX-FCM) was selected from the aforementioned methods and was compared to the SA-MIX-FCM scheme. We used the percent value of the normalized interquartile range (nIQR) to its median to evaluate the variability of all methods. 2D-FCM clustering is more effective than 1D-FCM clustering at the severe overload range only, but inferior for other ranges (where 1D-FCM provides suitable results). This complementary performance between the two methods allows MIX-FCM to improve results for all ranges. OP-MIX-FCM clustering error was 2.1 ± 2.3 %, compared with 10.3 ± 9.9 % and 7.0 ± 11.9 % from 1D- and 2D-FCM clustering, respectively. SA-MIX-FCM result was comparable to OP-MIX-FCM result, with both schemes showing ability to decrease overall nIQR by approximately 30 %. Our proposed 2D-FCM algorithm is not as superior to 1D-FCM as hypothesized. In contrast, our MIX-FCM method benefits from the best of both methods to obtain the highest segmentation accuracy at all ranges. Moreover, segmentation accuracy of the practical scheme (SA-MIX-FCM) is comparable to segmentation accuracy of the reference scheme (OP-MIX-FCM). Finally, we confirmed that segmentation is crucial to improving LIC assessments, especially at the severe iron overload range.

Background Intracranial hemorrhage (ICH) treatment requires a prompt diagnosis based on a CT scan evaluation by a radiologist. Manually analyzing the images is complex and time-consuming. Deep-learning techniques have been successfully applied to assist radiologists with this task. We further improved the detection and localization of ICH lesions without subtype classification using the active learning concept and the Probabilistic CAM (PCAM) heatmap. Methods The train set of the second stage 2019-RSNA ICH data set was randomly separated into the train (712,803 slices), validate (20,000), and test (20,000) data sets and interpolated from 512 into 256, 1024, and 2048 image resolutions. All resolutions were scaled to RGB channels according to their subdural, brain, and bone windows and used in a DenseNet-121 2D-CNN classifier with the PCAM technique for localization using ICH labels. The point closest-to-(0,1) corner approach was used to find a cutoff point of the training data set on each image resolution and applied the cutoff value to the validate and test data sets to calculate performance metrics. Two neuroradiologists reevaluated the mismatched data between the RSNA labels and prediction outcomes on the test data set and reassigned labels when the evaluations agreed with the prediction outcomes. Both radiologists also evaluated the usefulness of the PCAM heatmap to localize ICH lesions into three categories: not useful (Poor), functional (Acceptable), and valuable (Good) grades. The improvement in ICH detection after reassigned RSNA labels was measured by AUROC and AUPRC. McNemar’s test determined whether image resolution would have a similar detection outcome with a significant level at P < 0.05. Results The study found that detecting ICH at 512 and 1024 resolutions gave a comparable performance and was significantly higher than at 256 and 2048 resolutions. However, their AUROC and AUPRC were still in the same range of 0.98 and 0.93, respectively. After reevaluation, 55% (471 from 859 FP) and 51% (114 from 225 FN) of the test data set were relabeled. Furthermore, using the same cutoff value, the AUROC and AUPRC can be increased by 1.1% (0.984 vs. 0.995) and 5.3% (0.932 vs. 0.981), respectively. The PCAM heatmaps obtained a Good grade of around 86%, 37% and 0% at 2048, 1024, and 512 resolutions, respectively. Conclusion Image resolution has a minor effect on altering the ICH detection performance, while reassessing the RSNA labels can significantly improve the performance. PCAM heatmaps can better localize the boundary regions of the ICH lesion at higher resolutions, with the best result in 2048.