Comparative evaluation of uncertainty estimation and decomposition methods on liver segmentation.

PURPOSE: Deep neural networks need to be able to indicate error likelihood via reliable estimates of their predictive uncertainty when used in high-risk scenarios, such as medical decision support. This work contributes a systematic overview of state-of-the-art approaches for decomposing predictive uncertainty into aleatoric and epistemic components, and a comprehensive comparison for Bayesian neural networks (BNNs) between mutual information decomposition and the explicit modelling of both uncertainty types via an additional loss-attenuating neuron. METHODS: Experiments are performed in the context of liver segmentation in CT scans. The quality of the uncertainty decomposition in the resulting uncertainty maps is qualitatively evaluated, and quantitative behaviour of decomposed uncertainties is systematically compared for different experiment settings with varying training set sizes, label noise, and distribution shifts. RESULTS: Our results show the mutual information decomposition to robustly yield meaningful aleatoric and epistemic uncertainty estimates, while the activation of the loss-attenuating neuron appears noisier with non-trivial convergence properties. We found that the addition of a heteroscedastic neuron does not significantly improve segmentation performance or calibration, while slightly improving the quality of uncertainty estimates. CONCLUSIONS: Mutual information decomposition is simple to implement, has mathematically pleasing properties, and yields meaningful uncertainty estimates that behave as expected under controlled changes to our data set. The additional extension of BNNs with loss-attenuating neurons provides no improvement in terms of segmentation performance or calibration in our setting, but marginal benefits regarding the quality of decomposed uncertainties.

Suitability of DNN-based vessel segmentation for SIRT planning.

PURPOSE: The segmentation of the hepatic arteries (HA) is essential for state-of-the-art pre-interventional planning of selective internal radiation therapy (SIRT), a treatment option for malignant tumors in the liver. In SIRT a catheter is placed through the aorta into the tumor-feeding hepatic arteries, injecting small beads filled with radiation emitting material for local radioembolization. In this study, we evaluate the suitability of a deep neural network (DNN) based vessel segmentation for SIRT planning. METHODS: We applied our DNN-based HA segmentation on 36 contrast-enhanced computed tomography (CT) scans from the arterial contrast agent phase and rated its segmentation quality as well as the overall image quality. Additionally, we applied a traditional machine learning algorithm for HA segmentation as comparison to our deep learning (DL) approach. Moreover, we assessed by expert ratings whether the produced HA segmentations can be used for SIRT planning. RESULTS: The DL approach outperformed the traditional machine learning algorithm. The DL segmentation can be used for SIRT planning in [Formula: see text] of the cases, while the reference segmentations, which were manually created by experienced radiographers, are sufficient in [Formula: see text]. Seven DL cases cannot be used for SIRT planning while the corresponding reference segmentations are sufficient. However, there are two DL segmentations usable for SIRT, where the reference segmentations for the same cases were rated as insufficient. CONCLUSIONS: HA segmentation is a difficult and time-consuming task. DL-based methods have the potential to support and accelerate the pre-interventional planning of SIRT therapy.

Influence of Medial Patellofemoral Ligament Reconstruction on Patellofemoral Contact in Patients With Low-Flexion Patellar Instability: An MRI Study.

Background: Medial patellofemoral ligament (MPFL) reconstruction is a well-established procedure for the treatment of patients with patellofemoral instability (PFI) at low flexion angles (0°-30°). Little is known about the effect of MPFL surgery on patellofemoral cartilage contact area (CCA) during the first 30° of knee flexion. Purpose/Hypothesis: The purpose of this study was to investigate the effect of MPFL reconstruction on CCA using magnetic resonance imaging (MRI). We hypothesized that patients with PFI would have a lower CCA than patients with healthy knees and that CCA would increase after MPFL reconstruction over the course of low knee flexion. Study Design: Cohort study; Level of evidence, 2. Methods: In a prospective matched-paired cohort study, the CCA of 13 patients with low-flexion PFI was determined before and after MPFL reconstruction, and the data were compared with those of 13 healthy volunteers (controls). MRI was performed with the knee at 0°, 15°, and 30° of flexion in a custom-designed knee-positioning device. To suppress motion artifacts, motion correction was performed using a Moiré Phase Tracking system via a tracking marker attached to the patella. The CCA was calculated on the basis of semiautomatic cartilage and bone segmentation and registration. Results: The CCA (mean ± SD) at 0°, 15°, and 30° of flexion for the control participants was 1.38 ± 0.62, 1.91 ± 0.98, and 3.68 ± 0.92 cm2, respectively. In patients with PFI, the CCA at 0°, 15°, and 30° of flexion was 0.77 ± 0.49, 1.26 ± 0.60, and 2.89 ± 0.89 cm2 preoperatively and 1.65 ± 0.55, 1.97 ± 0.68, and 3.52 ± 0.57 cm2 postoperatively. Patients with PFI exhibited a significantly reduced preoperative CCA at all 3 flexion angles when compared with controls (P ≤ .045 for all). Postoperatively, there was a significant increase in CCA at 0° of flexion (P = .001), 15° of flexion (P = .019) and 30° of flexion (P = .026). There were no significant postoperative differences in CCA between patients with PFI and controls at any flexion angle. Conclusion: Patients with low-flexion patellar instability showed a significant reduction in patellofemoral CCA at 0°, 15°, and 30° of flexion. MPFL reconstruction increased the contact area significantly at all angles.

Comparative validation of machine learning algorithms for surgical workflow and skill analysis with the HeiChole benchmark.

PURPOSE: Surgical workflow and skill analysis are key technologies for the next generation of cognitive surgical assistance systems. These systems could increase the safety of the operation through context-sensitive warnings and semi-autonomous robotic assistance or improve training of surgeons via data-driven feedback. In surgical workflow analysis up to 91% average precision has been reported for phase recognition on an open data single-center video dataset. In this work we investigated the generalizability of phase recognition algorithms in a multicenter setting including more difficult recognition tasks such as surgical action and surgical skill. METHODS: To achieve this goal, a dataset with 33 laparoscopic cholecystectomy videos from three surgical centers with a total operation time of 22 h was created. Labels included framewise annotation of seven surgical phases with 250 phase transitions, 5514 occurences of four surgical actions, 6980 occurences of 21 surgical instruments from seven instrument categories and 495 skill classifications in five skill dimensions. The dataset was used in the 2019 international Endoscopic Vision challenge, sub-challenge for surgical workflow and skill analysis. Here, 12 research teams trained and submitted their machine learning algorithms for recognition of phase, action, instrument and/or skill assessment. RESULTS: F1-scores were achieved for phase recognition between 23.9% and 67.7% (n = 9 teams), for instrument presence detection between 38.5% and 63.8% (n = 8 teams), but for action recognition only between 21.8% and 23.3% (n = 5 teams). The average absolute error for skill assessment was 0.78 (n = 1 team). CONCLUSION: Surgical workflow and skill analysis are promising technologies to support the surgical team, but there is still room for improvement, as shown by our comparison of machine learning algorithms. This novel HeiChole benchmark can be used for comparable evaluation and validation of future work. In future studies, it is of utmost importance to create more open, high-quality datasets in order to allow the development of artificial intelligence and cognitive robotics in surgery.

Change in Descriptive Kinematic Parameters of Patients with Patellofemoral Instability When Compared to Individuals with Healthy Knees-A 3D MRI In Vivo Analysis.

BACKGROUND: Patellofemoral instability (PFI) leads to chronic knee pain, reduced performance and chondromalacia patellae with consecutive osteoarthritis. Therefore, determining the exact patellofemoral contact mechanism, as well as the factors leading to PFI, is of great importance. The present study compares in vivo patellofemoral kinematic parameters and the contact mechanism of volunteers with healthy knees and patients with low flexion patellofemoral instability (PFI). The study was performed with a high-resolution dynamic MRI. MATERIAL/METHODS: In a prospective cohort study, the patellar shift, patella rotation and the patellofemoral cartilage contact areas (CCA) of 17 patients with low flexion PFI were analyzed and compared with 17 healthy volunteers, matched via the TEA distance and sex, in unloaded and loaded conditions. MRI scans were carried out for 0°, 15° and 30° knee flexion in a custom-designed knee loading device. To suppress motion artifacts, motion correction was performed using a moiré phase tracking system with a tracking marker attached to the patella. The patellofemoral kinematic parameters and the CCA was calculated on the basis of semi-automated cartilage and bone segmentation and registrations. RESULTS: Patients with low flexion PFI showed a significant reduction in patellofemoral CCA for 0° (unloaded: p = 0.002, loaded: p = 0.004), 15° (unloaded: p = 0.014, loaded: p = 0.001) and 30° (unloaded: p = 0.008; loaded: p = 0.001) flexion compared to healthy subjects. Additionally, patients with PFI revealed a significantly increased patellar shift when compared to volunteers with healthy knees at 0° (unloaded: p = 0.033; loaded: p = 0.031), 15° (unloaded: p = 0.025; loaded: p = 0.014) and 30° flexion (unloaded: p = 0.030; loaded: p = 0.034) There were no significant differences for patella rotation between patients with PFI and the volunteers, except when, under load at 0° flexion, PFI patients showed increased patellar rotation (p = 0.005. The influence of quadriceps activation on the patellofemoral CCA is reduced in patients with low flexion PFI. CONCLUSION: Patients with PFI showed different patellofemoral kinematics at low flexion angles in both unloaded and loaded conditions compared to volunteers with healthy knees. Increased patellar shifts and decreased patellofemoral CCAs were observed in low flexion angles. The influence of the quadriceps muscle is diminished in patients with low flexion PFI. Therefore, the goal of patellofemoral stabilizing therapy should be to restore a physiologic contact mechanism and improve patellofemoral congruity for low flexion angles.

Individual Influence of Trochlear Dysplasia on Patellofemoral Kinematics after Isolated MPFL Reconstruction.

INTRODUCTION: The influence of the MPFL graft in cases of patella instability with dysplastic trochlea is a controversial topic. The effect of the MPFL reconstruction as single therapy is under investigation, especially with severely dysplastic trochlea (Dejour types C and D). The purpose of this study was to evaluate the impact of trochlear dysplasia on patellar kinematics in patients suffering from low flexion patellar instability under weight-bearing conditions after isolated MPFL reconstruction. MATERIAL AND METHODS: Thirteen patients were included in this study, among them were eight patients with mild dysplasia (Dejour type A and B) and five patients with severe dysplasia (Dejour type C and D). By performing a knee MRI with in situ loading, patella kinematics and the patellofemoral cartilage contact area could be measured under the activation of the quadriceps musculature in knee flexion angles of 0°, 15° and 30°. To mitigate MRI motion artefacts, prospective motion correction based on optical tracking was applied. Bone and cartilage segmentation were performed semi-automatically for further data analysis. Cartilage contact area (CCA) and patella tilt were the main outcome measures for this study. Pre- and post-surgery measures were compared for each group. RESULTS: Data visualized a trending lower patella tilt after MPFL graft installation in both groups and flexion angles of the knee. There were no significant changes in patella tilt at 0° (unloaded pre-surgery: 22.6 ± 15.2; post-surgery: 17.7 ± 14.3; p = 0.110) and unloaded 15° flexion (pre-surgery: 18.9 ± 12.7; post-surgery: 12.2 ± 13.0; p = 0.052) of the knee in patients with mild dysplasia, whereas in patients with severe dysplasia of the trochlea the results happened not to be significant in the same angles with loading of 5 kg (0° flexion pre-surgery: 34.4 ± 12.1; post-surgery: 31.2 ± 16.1; p = 0.5; 15° flexion pre-surgery: 33.3 ± 6.1; post-surgery: 23.4 ± 8.6; p = 0.068). CCA increased in every flexion angle and group, but significant increase was seen only between 0°-15° (unloaded and loaded) in mild dysplasia of the trochlea, where significant increase in Dejour type C and D group was seen with unloaded full extension of the knee (0° flexion) and 30° flexion (unloaded and loaded). CONCLUSION: This study proves a significant effect of the MPFL graft to cartilage contact area, as well as an improvement of the patella tilt in patients with mild dysplasia of the trochlea. Thus, the MPFL can be used as a single treatment for patient with Dejour type A and B dysplasia. However, in patients with severe dysplasia the MPFL graft alone does not significantly increase CCA.

Quantitative Analysis of Liver Disease Using MRI-Based Radiomic Features of the Liver and Spleen.

BACKGROUND: Radiomics extracts quantitative image features to identify biomarkers for characterizing disease. Our aim was to characterize the ability of radiomic features extracted from magnetic resonance (MR) imaging of the liver and spleen to detect cirrhosis by comparing features from patients with cirrhosis to those without cirrhosis. METHODS: This retrospective study compared MR-derived radiomic features between patients with cirrhosis undergoing hepatocellular carcinoma screening and patients without cirrhosis undergoing intraductal papillary mucinous neoplasm surveillance between 2015 and 2018 using the same imaging protocol. Secondary analyses stratified the cirrhosis cohort by liver disease severity using clinical compensation/decompensation and Model for End-Stage Liver Disease (MELD). RESULTS: Of 167 patients, 90 had cirrhosis with 68.9% compensated and median MELD 8. Combined liver and spleen radiomic features generated an AUC 0.94 for detecting cirrhosis, with shape and texture components contributing more than size. Discrimination of cirrhosis remained high after stratification by liver disease severity. CONCLUSIONS: MR-based liver and spleen radiomic features had high accuracy in identifying cirrhosis, after stratification by clinical compensation/decompensation and MELD. Shape and texture features performed better than size features. These findings will inform radiomic-based applications for cirrhosis diagnosis and severity.

Improving automatic liver tumor segmentation in late-phase MRI using multi-model training and 3D convolutional neural networks.

Automatic liver tumor segmentation can facilitate the planning of liver interventions. For diagnosis of hepatocellular carcinoma, dynamic contrast-enhanced MRI (DCE-MRI) can yield a higher sensitivity than contrast-enhanced CT. However, most studies on automatic liver lesion segmentation have focused on CT. In this study, we present a deep learning-based approach for liver tumor segmentation in the late hepatocellular phase of DCE-MRI, using an anisotropic 3D U-Net architecture and a multi-model training strategy. The 3D architecture improves the segmentation performance compared to a previous study using a 2D U-Net (mean Dice 0.70 vs. 0.65). A further significant improvement is achieved by a multi-model training approach (0.74), which is close to the inter-rater agreement (0.78). A qualitative expert rating of the automatically generated contours confirms the benefit of the multi-model training strategy, with 66 % of contours rated as good or very good, compared to only 43 % when performing a single training. The lesion detection performance with a mean F1-score of 0.59 is inferior to human raters (0.76). Overall, this study shows that correctly detected liver lesions in late-phase DCE-MRI data can be automatically segmented with high accuracy, but the detection, in particular of smaller lesions, can still be improved.

Reconstruction of dental roots for implant planning purposes: a feasibility study.

PURPOSE: Modern virtual implant planning is a time-consuming procedure, requiring a careful assessment of prosthetic and anatomical factors within a three-dimensional dataset. In order to facilitate the planning process and provide additional information, this study examines a statistical shape model (SSM) to compute the course of dental roots based on a surface scan. MATERIAL AND METHODS: Plaster models of orthognathic patients were scanned and superimposed with three-dimensional data of a cone-beam computer tomography (CBCT). Based on the open-source software "R", including the packages Morpho, mesheR, Rvcg and RvtkStatismo, an SSM was generated to estimate the tooth axes. The accuracy of the calculated tooth axes was determined using a leave-one-out cross-validation. The deviation of tooth axis prediction in terms of angle or horizontal shift is described with mean and standard deviation. The planning dataset of an implant surgery patient was additionally analyzed using the SSM. RESULTS: 71 datasets were included in this study. The mean angle between the estimated tooth-axis and the actual tooth-axis was 7.5 ± 4.3° in the upper jaw and 6.7 ± 3.8° in the lower jaw. The horizontal deviation between the tooth axis and estimated axis was 1.3 ± 0.8 mm close to the cementoenamel junction, and 0.7 ± 0.5 mm in the apical third of the root. Results for models with one missing tooth did not differ significantly. In the clinical dataset, the SSM could give a reasonable aid for implant positioning. CONCLUSIONS: With the presented SSM, the approximate course of dental roots can be predicted based on a surface scan. There was no difference in predicting the tooth axis of existent or missing teeth. In clinical context, the estimation of tooth axes of missing teeth could serve as a reference for implant positioning. However, a higher number of training data must be achieved to obtain increasing accuracy.

MRI-based radiomic feature analysis of end-stage liver disease for severity stratification.

PURPOSE: We aimed to develop a predictive model of disease severity for cirrhosis using MRI-derived radiomic features of the liver and spleen and compared it to the existing disease severity metrics of MELD score and clinical decompensation. The MELD score is compiled solely by blood parameters, and so far, it was not investigated if extracted image-based features have the potential to reflect severity to potentially complement the calculated score. METHODS: This was a retrospective study of eligible patients with cirrhosis ([Formula: see text]) who underwent a contrast-enhanced MR screening protocol for hepatocellular carcinoma (HCC) screening at a tertiary academic center from 2015 to 2018. Radiomic feature analyses were used to train four prediction models for assessing the patient's condition at time of scan: MELD score, MELD score [Formula: see text] 9 (median score of the cohort), MELD score [Formula: see text] 15 (the inflection between the risk and benefit of transplant), and clinical decompensation. Liver and spleen segmentations were used for feature extraction, followed by cross-validated random forest classification. RESULTS: Radiomic features of the liver and spleen were most predictive of clinical decompensation (AUC 0.84), which the MELD score could predict with an AUC of 0.78. Using liver or spleen features alone had slightly lower discrimination ability (AUC of 0.82 for liver and AUC of 0.78 for spleen features only), although this was not statistically significant on our cohort. When radiomic prediction models were trained to predict continuous MELD scores, there was poor correlation. When stratifying risk by splitting our cohort at the median MELD 9 or at MELD 15, our models achieved AUCs of 0.78 or 0.66, respectively. CONCLUSIONS: We demonstrated that MRI-based radiomic features of the liver and spleen have the potential to predict the severity of liver cirrhosis, using decompensation or MELD status as imperfect surrogate measures for disease severity.

Anisotropic 3D Multi-Stream CNN for Accurate Prostate Segmentation from Multi-Planar MRI.

BACKGROUND AND OBJECTIVE: Accurate and reliable segmentation of the prostate gland in MR images can support the clinical assessment of prostate cancer, as well as the planning and monitoring of focal and loco-regional therapeutic interventions. Despite the availability of multi-planar MR scans due to standardized protocols, the majority of segmentation approaches presented in the literature consider the axial scans only. In this work, we investigate whether a neural network processing anisotropic multi-planar images could work in the context of a semantic segmentation task, and if so, how this additional information would improve the segmentation quality. METHODS: We propose an anisotropic 3D multi-stream CNN architecture, which processes additional scan directions to produce a high-resolution isotropic prostate segmentation. We investigate two variants of our architecture, which work on two (dual-plane) and three (triple-plane) image orientations, respectively. The influence of additional information used by these models is evaluated by comparing them with a single-plane baseline processing only axial images. To realize a fair comparison, we employ a hyperparameter optimization strategy to select optimal configurations for the individual approaches. RESULTS: Training and evaluation on two datasets spanning multiple sites show statistical significant improvement over the plain axial segmentation (p<0.05 on the Dice similarity coefficient). The improvement can be observed especially at the base (0.898 single-plane vs. 0.906 triple-plane) and apex (0.888 single-plane vs. 0.901 dual-plane). CONCLUSION: This study indicates that models employing two or three scan directions are superior to plain axial segmentation. The knowledge of precise boundaries of the prostate is crucial for the conservation of risk structures. Thus, the proposed models have the potential to improve the outcome of prostate cancer diagnosis and therapies.

Fluoroscopic guided tunnel placement during medial patellofemoral ligament reconstruction is not accurate in patients with severe trochlear dysplasia.

PURPOSE: Accurate femoral tunnel placement is of great importance during medial patellofemoral ligament (MPFL) reconstruction. Purpose of the present study was to investigate the influence of trochlear dysplasia on the accuracy of fluoroscopic guided femoral tunnel placement. METHODS: CT-Scans of 30 knees (five with regular shaped trochlea, 10 with a Type A and five each with a Type B, C, or D trochlear dysplasia) were imported into the image analysis platform MeVisLab. A 3D Bone Volume Rendering (VR) and a virtual lateral radiograph was created. The anatomic femoral MPFL insertion was identified on the 3D VR. On virtual lateral radiographs, the MPFL insertion was identified based on landmarks described by Schöttle et al. using three different perspectives: Best possible overlap of the femoral condyles (BC) and a tangent along posterior border of the posterior femoral cortex (pBC); a tangent along the anterior border of the posterior cortex (aBC); and best possible overlap of the distal part of the posterior femoral cortex (BF). Distances between the anatomic attachment and radiographically obtained insertions were measured on the 3D VR and compared according to the type of trochlear dysplasia. RESULTS: Significantly lower accuracy of fluoroscopy guided tunnel placement in MPFL reconstruction was found in knees with Type C and D dysplasia. This effect was observed irrespectively from the radiologic perspective (pBC, aBC, and FC). In the pBC view (highest accuracy), the mean distance from the centre of the anatomic MPFL attachment to the radiographically defined location was 4.3 mm in knees without trochlear dysplasia and increased to 4.8 mm in knees with Type A dysplasia, 3.8 mm in knees with Type B dysplasia, 6.7 mm (p < 0.001) in knees with Type C dysplasia, and 7.3 mm (p < 0.001) in knees with Type D dysplasia. CONCLUSION: Radiographic landmark-based femoral tunnel placement in the pBC view provides highest accuracy in knees with a normal shaped trochlea or low grade trochlear dysplasia. In patients with severe dysplasia, fluoroscopy guided tunnel placement has a low accuracy, exceeding a critical threshold of 5 mm distance to the anatomic MPFL insertion irrespective of the radiographic perspective. In these patients, utilization of anatomic landmarks may be beneficial. LEVEL OF EVIDENCE: IV.

Reducing inter-observer variability and interaction time of MR liver volumetry by combining automatic CNN-based liver segmentation and manual corrections.

PURPOSE: To compare manual corrections of liver masks produced by a fully automatic segmentation method based on convolutional neural networks (CNN) with manual routine segmentations in MR images in terms of inter-observer variability and interaction time. METHODS: For testing, patient's precise reference segmentations that fulfill the quality requirements for liver surgery were manually created. One radiologist and two radiology residents were asked to provide manual routine segmentations. We used our automatic segmentation method Liver-Net to produce liver masks for the test cases and asked a radiologist assistant and one further resident to correct the automatic results. All observers were asked to measure their interaction time. Both manual routine and corrected segmentations were compared with the reference annotations. RESULTS: The manual routine segmentations achieved a mean Dice index of 0.95 and a mean relative error (RVE) of 4.7%. The quality of liver masks produced by the Liver-Net was on average 0.95 Dice and 4.5% RVE. Liver masks resulting from manual corrections of automatically generated segmentations compared to routine results led to a significantly lower inter-observer variability (mean per case absolute RVE difference across observers 0.69%) when compared to manual routine ones (2.75%). The mean interaction time was 2 min for manual corrections and 10 min for manual routine segmentations. CONCLUSIONS: The quality of automatic liver segmentations is on par with those from manual routines. Using automatic liver masks in the clinical workflow could lead to a reduction of segmentation time and a more consistent liver volume estimation across different observers.

Evaluation of deep learning methods for parotid gland segmentation from CT images.

The segmentation of organs at risk is a crucial and time-consuming step in radiotherapy planning. Good automatic methods can significantly reduce the time clinicians have to spend on this task. Due to its variability in shape and low contrast to surrounding structures, segmenting the parotid gland is challenging. Motivated by the recent success of deep learning, we study the use of two-dimensional (2-D), 2-D ensemble, and three-dimensional (3-D) U-Nets for segmentation. The mean Dice similarity to ground truth is ∼ 0.83 for all three models. A patch-based approach for class balancing seems promising for false-positive reduction. The 2-D ensemble and 3-D U-Net are applied to the test data of the 2015 MICCAI challenge on head and neck autosegmentation. Both deep learning methods generalize well onto independent data (Dice 0.865 and 0.88) and are superior to a selection of model- and atlas-based methods with respect to the Dice coefficient. Since appropriate reference annotations are essential for training but often difficult and expensive to obtain, it is important to know how many samples are needed for training. We evaluate the performance after training with different-sized training sets and observe no significant increase in the Dice coefficient for more than 250 training cases.

Automatic liver tumor segmentation in CT with fully convolutional neural networks and object-based postprocessing.

Automatic liver tumor segmentation would have a big impact on liver therapy planning procedures and follow-up assessment, thanks to standardization and incorporation of full volumetric information. In this work, we develop a fully automatic method for liver tumor segmentation in CT images based on a 2D fully convolutional neural network with an object-based postprocessing step. We describe our experiments on the LiTS challenge training data set and evaluate segmentation and detection performance. Our proposed design cascading two models working on voxel- and object-level allowed for a significant reduction of false positive findings by 85% when compared with the raw neural network output. In comparison with the human performance, our approach achieves a similar segmentation quality for detected tumors (mean Dice 0.69 vs. 0.72), but is inferior in the detection performance (recall 63% vs. 92%). Finally, we describe how we participated in the LiTS challenge and achieved state-of-the-art performance.

Segmentation-based partial volume correction for volume estimation of solid lesions in CT.

In oncological chemotherapy monitoring, the change of a tumor's size is an important criterion for assessing cancer therapeutics. Measuring the volume of a tumor requires its delineation in 3-D. This is called segmentation, which is an intensively studied problem in medical image processing. However, simply counting the voxels within a binary segmentation result can lead to significant differences in the volume, if the lesion has been segmented slightly differently by various segmentation procedures or in different scans, for example due to the limited spatial resolution of computed tomography (CT) or partial volume effects. This variability limits the sensitivity of size measurements and thus of therapy response assessments and it can even lead to misclassifications. We present a fast, generic algorithm for measuring the volume of solid, compact tumors in CT that considers partial volume effects at the border of a given segmentation result. The algorithm is an extension of the segmentation-based partial volume analysis proposed by Kuhnigk for the volumetry of solid lung lesions , such that it can be applied to inhomogeneous lesions and lesions with inhomogeneous surroundings. Our generalized segmentation-based partial volume correction is based on a spatial subdivision of the segmentation result, from which the fraction of tumor for each voxel is computed. It has been evaluated on phantom data, 1516 lesion segmentation pairs (lung nodules, liver metastases and lymph nodes) as well as 1851 lung nodules from the LIDC-IDRI database. The evaluations of our algorithm show a more accurate estimation of the real volume and its ability to reduce inter- and intra-observer variability significantly for each entity. Overall, the variability (interquartile range) for phantom data is reduced by 49% ( p ≪ 0.001) and the variability between different readers is reduced by 28% ( p ≪ 0.001). The average computation time is 0.2 s.

On the evaluation of segmentation editing tools.

Efficient segmentation editing tools are important components in the segmentation process, as no automatic methods exist that always generate sufficient results. Evaluating segmentation editing algorithms is challenging, because their quality depends on the user's subjective impression. So far, no established methods for an objective, comprehensive evaluation of such tools exist and, particularly, intermediate segmentation results are not taken into account. We discuss the evaluation of editing algorithms in the context of tumor segmentation in computed tomography. We propose a rating scheme to qualitatively measure the accuracy and efficiency of editing tools in user studies. In order to objectively summarize the overall quality, we propose two scores based on the subjective rating and the quantified segmentation quality over time. Finally, a simulation-based evaluation approach is discussed, which allows a more reproducible evaluation without the need for human input. This automated evaluation complements user studies, allowing a more convincing evaluation, particularly during development, where frequent user studies are not possible. The proposed methods have been used to evaluate two dedicated editing algorithms on 131 representative tumor segmentations. We show how the comparison of editing algorithms benefits from the proposed methods. Our results also show the correlation of the suggested quality score with the qualitative ratings.