Auto-segmentation Algorithm Research Articles

Autosegmentation based on different-sized training datasets of consistently-curated volumes and impact on rectal contours in prostate cancer radiation therapy.

Background and purposeAutosegmentation techniques are emerging as time-saving means for radiation therapy (RT) contouring, but the understanding of their performance on different datasets is limited. The aim of this study was to determine agreement between rectal volumes by an existing autosegmentation algorithm and manually-delineated rectal volumes in prostate cancer RT. We also investigated contour quality by different-sized training datasets and consistently-curated volumes for retrained versions of this same algorithm.Materials and methodsSingle-institutional data from 624 prostate cancer patients treated to 50–70 Gy were used. Manually-delineated clinical rectal volumes (clinical) and consistently-curated volumes recontoured to one anatomical guideline (reference) were compared to autocontoured volumes by a commercial autosegmentation tool based on deep-learning (v1; n = 891, multiple-institutional data) and retrained versions using subsets of the curated volumes (v32/64/128/256; n = 32/64/128/256). Evaluations included dose-volume histogram metrics, Dice similarity coefficients, and Hausdorff distances; differences between groups were quantified using parametric or non-parametric hypothesis testing.ResultsVolumes by v1-256 (76–78 cm3) were larger than reference (75 cm3) and clinical (76 cm3). Mean doses by v1-256 (24.2–25.2 Gy) were closer to reference (24.2 Gy) than to clinical (23.8 Gy). Maximum doses were similar for all volumes (65.7–66.0 Gy). Dice for v1-256 and reference (0.87–0.89) were higher than for v1-256 and clinical (0.86–0.87) with corresponding Hausdorff comparisons including reference smaller than comparisons including clinical (5–6 mm vs. 7–8 mm).ConclusionUsing small single-institutional RT datasets with consistently-defined rectal volumes when training autosegmentation algorithms created contours of similar quality as the same algorithm trained on large multi-institutional datasets.

Development of a Deep Learning-Based Auto-Segmentation of Organs at Risk for Head and Neck Radiotherapy Planning

<h3>Purpose/Objective(s)</h3> Segmentation of head and neck (HN) organs at risk (OARs) is a laborious process. Here we introduce and validate a newly developed deep-learning-based auto-segmentation program and compare with a commercially available system, and the same system trained on internal data. <h3>Materials/Methods</h3> A total of 864 previously treated HN cancer patients were available to train and evaluate a prototype deep learning-based normal tissue 3D auto-segmentation algorithm. The algorithm is based on a fully convolutional network with U-Net and V-net features combined as the backbone of the network. A Dice loss function with the Adam optimizer was used in training the models with between 150-500 patients used per model. The OARs were delineated by a single experienced physician (gold data). A subset of 75 cases was withheld from training and used for validation. On those, we generated new OAR sets with three different deep-learning models and compared to the gold data: A) the prototype model trained with gold data, B) a commercial software package trained with the gold data (n=213), and C) the same commercial software with the model trained at another institution (n=589). The agreement between the gold data and auto-segmented structures was evaluated with Dice similarity coefficient (DSC) and voxel-penalty metric that penalizes each missing or extra pixel as a function of distance with forgiveness threshold distance. An ANOVA test with post hoc pair-wise analysis was performed to assess the differences in those metrics. The auto-segmented contours were also qualitatively evaluated by the physician on a scale of 0-5. <h3>Results</h3> The average DSC and voxel penalty metric scores for algorithms A, B, and C across all OARs in the 75 evaluation cases were 0.80/77.68, 0.74/62.75, and 0.66/45.26, respectively. The difference in mean DSC scores was statistically significant (p<0.05) for all 11 OARs where the data for all three algorithms were available. The A/B difference was significant in 6 OARs. Algorithm A scored the highest DSC and voxel penalty metric score in all OARs except for the pharyngeal constrictors. All OARs except for the pharyngeal constrictors showed DSC≥0.7 with algorithm A. For three structures the mean DSC was significantly different between the same algorithm trained at different institutions(B/C). From the qualitative evaluation by a blinded expert, 51 structures (20.2%) of model A were clinically acceptable without edits. The percentages of ‘clinically useful' scores were largest in model A (95.2%) followed by model B (88.0%) and C (80.6%). <h3>Conclusion</h3> The prototype algorithm had improved performance compared to a commercial algorithm, even when trained on data from the same institution. Auto segmentation results can differ significantly when the same algorithm is trained on data from different institutions.

A Novel and Automated Approach to Classify Radiation Induced Lung Tissue Damage on CT Scans.

Simple SummaryRadiation-induced lung damage (RILD) is a common side effect of treating lung cancer with radiotherapy (RT). RILD is visible on CT imaging, and its radiological appearance can vary dramatically from patient to patient as well as across different sub-regions of the lung and treatment volumes. A classification system for RILD able to differentiate radiological damage on a local level would allow us to better understand the underlying patterns of RILD, see how they change over time post irradiation, and link it with clinical outcomes. In this work we propose a five-class morphological lung tissue classification system that can describe parenchymal tissue changes at the voxel level. The classifier was implemented in a fully automated manner using an optimised deep-learning method, then trained and tested using data acquired through a multi-centre clinical trial. The proposed method performed well on an unseen testing dataset. The automated segmentation achieved considerable overlap with manual segmentations (ranging between 26% and 98% for the five classes) and was graded as acceptable by a clinical expert in 88% of cases. This demonstrates it to be suitable for application on a large dataset to help uncover different patterns of changes in the population.Radiation-induced lung damage (RILD) is a common side effect of radiotherapy (RT). The ability to automatically segment, classify, and quantify different types of lung parenchymal change is essential to uncover underlying patterns of RILD and their evolution over time. A RILD dedicated tissue classification system was developed to describe lung parenchymal tissue changes on a voxel-wise level. The classification system was automated for segmentation of five lung tissue classes on computed tomography (CT) scans that described incrementally increasing tissue density, ranging from normal lung (Class 1) to consolidation (Class 5). For ground truth data generation, we employed a two-stage data annotation approach, akin to active learning. Manual segmentation was used to train a stage one auto-segmentation method. These results were manually refined and used to train the stage two auto-segmentation algorithm. The stage two auto-segmentation algorithm was an ensemble of six 2D Unets using different loss functions and numbers of input channels. The development dataset used in this study consisted of 40 cases, each with a pre-radiotherapy, 3-, 6-, 12-, and 24-month follow-up CT scans (n = 200 CT scans). The method was assessed on a hold-out test dataset of 6 cases (n = 30 CT scans). The global Dice score coefficients (DSC) achieved for each tissue class were: Class (1) 99% and 98%, Class (2) 71% and 44%, Class (3) 56% and 26%, Class (4) 79% and 47%, and Class (5) 96% and 92%, for development and test subsets, respectively. The lowest values for the test subsets were caused by imaging artefacts or reflected subgroups that occurred infrequently and with smaller overall parenchymal volumes. We performed qualitative evaluation on the test dataset presenting manual and auto-segmentation to a blinded independent radiologist to rate them as ‘acceptable’, ‘minor disagreement’ or ‘major disagreement’. The auto-segmentation ratings were similar to the manual segmentation, both having approximately 90% of cases rated as acceptable. The proposed framework for auto-segmentation of different lung tissue classes produces acceptable results in the majority of cases and has the potential to facilitate future large studies of RILD.

Technical note: Evaluation of a V-Net autosegmentation algorithm for pediatric CT scans: Performance, generalizability, and application to patient-specific CT dosimetry.

This study developed and evaluated a fully convolutional network (FCN) for pediatric CT organ segmentation and investigated the generalizability of the FCN across image heterogeneities such as CT scanner model protocols and patient age. We also evaluated the autosegmentation models as part of a software tool for patient-specific CT dose estimation. A collection of 359 pediatric CT datasets with expert organ contours were used for model development and evaluation. Autosegmentation models were trained for each organ using a modified FCN 3D V-Net. An independent test set of 60 patients was withheld for testing. To evaluate the impact of CT scanner model protocol and patient age heterogeneities, separate models were trained using a subset of scanner model protocols and pediatric age groups. Train and test sets were split to answer questions about the generalizability of pediatric FCN autosegmentation models to unseen age groups and scanner model protocols, as well as the merit of scanner model protocol or age-group-specific models. Finally, the organ contours resulting from the autosegmentation models were applied to patient-specific dose maps to evaluate the impact of segmentation errors on organ doseestimation. Results demonstrate that the autosegmentation models generalize to CT scanner acquisition and reconstruction methods which were not present in the training dataset. While models are not equally generalizable across age groups, age-group-specific models do not hold any advantage over combining heterogeneous age groups into a single training set. Dice similarity coefficient (DSC) and mean surface distance results are presented for 19 organ structures, for example, median DSC of 0.52 (duodenum), 0.74 (pancreas), 0.92 (stomach), and 0.96 (heart). The FCN models achieve a mean dose error within 5% of expert segmentations for all 19 organs except for the spinal canal, where the mean error was 6.31%. Overall, these results are promising for the adoption of FCN autosegmentation models for pediatric CT, including applications for patient-specific CT doseestimation.

Development of Simplified Auto-Segmentable Functional Cardiac Atlas

PurposeThere is increasing evidence that radiation doses to cardiac substructures are associated with cardiac adverse events. Manual delineation of cardiac substructures is time-consuming, and auto-segmentation of cardiac substructure atlases has consequently been evaluated. However, proper automatic delineation of small substructures, such as the left anterior descending coronary artery, is challenging, and auto-segmentation of cardiac conduction system substructures has never been evaluated, despite multiple reports of radiation-induced arrhythmia after thoracic irradiations. The aim of this study was to propose and evaluate a simplified auto-segmentable functional cardiac atlas. Methods and MaterialsWe created a cardiac substructure atlas based on 20 computed tomography scans from patients with breast cancer comprising the 4 cardiac cavities, a high-risk cardiac zone as a left anterior descending coronary artery surrogate, and the 2 cardiac conduction nodes. Automatic delineation of this atlas by an atlas-based auto-segmentation algorithm was evaluated on a validation data set, consisting of 20 additional computed tomography scans. Dice similarity coefficients were used to evaluate the concordance level between the manual and the automatic contours; a dosimetric comparison between mean and maximum doses to the manual and to the auto-segmented substructures was additionally performed, based on intensity modulated radiation therapy treatment plans of the patients of the validation set. ResultsAverage dice similarity coefficient values were 0.78 for the 4 cardiac cavities, 0.65 for the high-risk cardiac zones, 0.56 for the sinoatrial node, and 0.15 for the atrioventricular node. Compared with manual contours, auto-segmented substructures were slightly smaller but the dosimetric parameters were similar. ConclusionsWe proposed a simplified functional cardiac atlas that included the cardiac conduction system and circumvented coronary delineation difficulties by using a surrogate high-risk cardiac zone. Most cardiac substructures were associated with acceptable atlas-based auto-segmentation properties. Such an atlas could be used for epidemiologic studies and for clinical practice.

Benchmarking Automatic Segmentation Algorithms Against Human Interobserver Variability of Prostate and Organs at Risk Delineation on Prostate MRI

<h3>Purpose/Objective(s)</h3> The objective of this study was to benchmark custom automatic segmentation software against human interobserver variability of prostate and OAR segmentation on 3D MRIs for prostate radiotherapy. We hypothesized that automatic segmentation algorithms can exhibit consistency comparable to human readers, including radiation oncologists. <h3>Materials/Methods</h3> Twenty-five patients underwent MRI-based treatment planning and quality assessment for low-dose-rate prostate brachytherapy (LDRPBT). All 25 patients received both treatment planning MRIs and postimplant quality assessment MRIs. The MRIs were acquired on 3T scanners with flexible external array coils and no endorectal coil. The preimplant MRIs were acquired with a 3D T2-weighted TSE sequence. The postimplant MRIs were acquired with a 3D T2/T1-weighted fully balanced SSFP sequence. Five human observers involved in the LDRPBT workflow, including 2 radiation oncologists, delineated the prostate and 4 OARs on the 50 MRIs. The OARs included the external urinary sphincter (EUS), seminal vesicles (SV), rectum, and bladder. Human interobserver variability (IoV) amongst the 5 human readers was quantified in a human observer study. Similarity metrics, including precision (P), recall (R), Jaccard index (JI), and Matthew's correlation coefficient (MCC), were quantified between unique observer pairs for all 5 organ segmentation masks. A custom deep learning-based autosegmentation algorithm was used to segment the prostate and 4 OARs simultaneously on the same 50 MRIs that were annotated by the human readers. P, R, JI, and MCC were evaluated at different probability thresholds. The comparisons were made against reference segmentation masks of the 5 organs computed from the 5 human segmentation masks using the simultaneous truth and performance level estimation algorithm. The human IoV similarity metrics were compared to the computer similarity metrics at the point of maximum MCC using two-tailed nonparametric Wilcoxon rank sum tests. <h3>Results</h3> For the preimplant MRIs, the computer similarity metrics (C) were lower than the human IoV similarity metrics (H) for all organs except the EUS (MCC, C vs. H: prostate – 0.868 vs. 0.903, <i>P</i> < 0.0001; EUS – 0.610 vs. 0.552, <i>P</i> = 0.1690; SV – 0.678 vs. 0.750, <i>P</i> = 0.0094; rectum – 0.790 vs. 0.886, <i>P</i> < 0.0001; bladder – 0.879 vs. 0.941, <i>P</i> < 0.0001). However, the opposite was observed for the postimplant MRIs; C was significantly higher than H for all 5 organs (MCC, C vs. H: prostate – 0.925 vs. 0.892, <i>P</i> < 0.0001; EUS – 0.743 vs. 0.620, <i>P</i> = 0.0003; SV – 0.849 vs. 0.760, <i>P</i> < 0.0001; rectum – 0.925 vs. 0.876, <i>P</i> < 0.0001; bladder – 0.972 vs. 0.938, <i>P</i> < 0.0001). <h3>Conclusion</h3> The autosegmentation algorithm produced high quality segmentation masks for both the preimplant and postimplant MRIs. The human readers were more consistent than the autosegmentation algorithm on the T2-weighted preimplant MRIs, but the autosegmentation algorithm was more consistent than the human readers on the postimplant MRIs.

BSEResU-Net: An attention-based before-activation residual U-Net for retinal vessel segmentation

Background and objectivesRetinal vessels are a major feature used for the physician to diagnose many retinal diseases, such as cardiovascular disease and Glaucoma. Therefore, the designing of an auto-segmentation algorithm for retinal vessel draw great attention in medical field. Recently, deep learning methods, especially convolutional neural networks (CNNs) show extraordinary potential for the task of vessel segmentation. However, most of the deep learning methods only take advantage of the shallow networks with a traditional cross-entropy objective, which becomes the main obstacle to further improve the performance on a task that is imbalanced. We therefore propose a new type of residual U-Net called Before-activation Squeeze-and-Excitation ResU-Net (BSEResu-Net) to tackle the aforementioned issues. MethodsOur BSEResU-Net can be viewed as an encoder/decoder framework that constructed by Before-activation Squeeze-and-Excitation blocks (BSE Blocks). In comparison to the current existing CNN structures, we utilize a new type of residual block structure, namely BSE block, in which the attention mechanism is combined with skip connection to boost the performance. What's more, the network could consistently gain accuracy from the increasing depth as we incorporate more residual blocks, attributing to the dropblock mechanism used in BSE blocks. A joint loss function which is based on the dice and cross-entropy loss functions is also introduced to achieve more balanced segmentation between the vessel and non-vessel pixels. ResultsThe proposed BSEResU-Net is evaluated on the publicly available DRIVE, STARE and HRF datasets. It achieves the F1-score of 0.8324, 0.8368 and 0.8237 on DRIVE, STARE and HRF dataset, respectively. Experimental results show that the proposed BSEResU-Net outperforms current state-of-the-art algorithms. ConclusionsThe proposed algorithm utilizes a new type of residual blocks called BSE residual blocks for vessel segmentation. Together with a joint loss function, it shows outstanding performance both on low and high-resolution fundus images.

Abstract PS15-15: Atlas-based auto-segmentation of a high risk cardiac zone on non-contrast computed tomography (CT) scans for indirect optimization of left anterior descending coronary artery (LADCA) dosimetry for breast radiotherapy

Abstract Background : Breast radiotherapy is associated with an increased cardiotoxicity risk which could be decreased by taking into account the left anterior descending coronary artery (LADCA) during treatment planning. However, LADCA manual contouring is time-consuming, complex and poorly reproducible on non-contrast simulation computed tomography (CT) scans where it is often hardly noticeable. Alternatively, auto-segmentation algorithms have been proposed for cardiac substructure contouring but they are usually unreliable for LADCA automatic delineation in daily practice. Purpose : The aim of this study was to implement and to evaluate auto-segmentation of a “high risk cardiac zone” which would be a reproducible LADCA surrogate and which would be reliably delineated by atlas-based algorithms for dosimetric purpose when planning breast radiotherapy on non-contrast CT scans. Materials/methods : 40 breast cancer patients treated with intensity modulated radiation therapy were randomly selected from our institutional database. “High risk cardiac zones” (HRCZ) were defined as segments of the anterior cardiac wall centered around the inter-ventricular groove from top to bottom (where the LADCA anatomically lies), with a constant 1cm-thickness and a symmetrical lateral margin on both sides of the groove. For each patient, eight HRCZ were contoured, differing by their width, ranging between 1 cm and 8 cm, in steps of 1 cm. Each contour was validated by a staff of three radiation oncologists. An atlas was constituted using the HRCZ contours of 20 patients and implemented in the “Workflow Box” (Mirada Medical) atlas-based auto-segmentation (ABAS) software. The ABAS algorithm delineated the HRCZ on the 20 remaining patients and the auto-contouring performances were evaluated by comparison with the manual HRCZ contours (defined as the reference contours), using dice-similarity coefficients (DSC) and Jaccard indexes, as a function of the HRCZ width. Results : Performances of the atlas-based auto-segmentation algorithm for HRCZ delineation are reported in table 1, as a function of HRCZ width. Auto-segmentation performance and reproducibility improved with HRCZ width, as evidenced by increased DSC and Jaccard indexes and lower relative standard variations; in particular, for a HRCZ width larger than 4 cm, HRCZ automatic segmentation was constantly satisfactory (DSC &amp;gt; 0.6; relative standard variation &amp;lt; 20%). Conclusion : We implemented and evaluated auto-segmentation of high risk cardiac zones (HRCZ) on non-contrast CT scans, defined as segments of the anterior cardiac wall geometrically centered around the inter-ventricular groove from top to bottom (where LADCA anatomically lies) with a constant thickness of 1cm and a total width ranging between 1 cm and 8 cm. For HRCZ width larger than 4cm, auto-segmented HRCZ contours were constantly reliable and may be used as LADCA surrogate for heart dosimetric optimization during breast radiotherapy planning on non-contrast CT scans. Table 1: performance of atlas-based auto-segmentation of high risk cardiac zones (HRCZ)HRCZ width (cm)DSC (mean)DSC SDDSC RSDJaccard Index (mean)Jaccard SDJaccard RSD10.3000.11563.35%0.1870.16254.04%20.5370.12533.14%0.3790.14026.05%30.6220.10623.06%0.4590.10416.69%40.6590.09919.95%0.4980.09314.06%50.6740.10219.81%0.5150.09213.67%60.6820.10419.77%0.5240.09213.43%70.6870.10419.67%0.5300.09113.27%80.6930.10519.57%0.5370.09213.23%Table 1: performance of atlas-based auto-segmentation of high risk cardiac zones (HRCZ), a LADCA surrogate, as a function of HRCZ width. DSC: Dice Similarity coefficient; sd : standard deviation; rsd : relative standard deviation. Citation Format: Pierre Loap, Nicolas Tkatchenko, Eliot Nicolas, Youlia Kirova. Atlas-based auto-segmentation of a high risk cardiac zone on non-contrast computed tomography (CT) scans for indirect optimization of left anterior descending coronary artery (LADCA) dosimetry for breast radiotherapy [abstract]. In: Proceedings of the 2020 San Antonio Breast Cancer Virtual Symposium; 2020 Dec 8-11; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2021;81(4 Suppl):Abstract nr PS15-15.

Abstract PS15-16: Evaluation of workflow box mirada software for automatic delineation of left anterior descending coronary artery (LADCA) in order reduce cardiac toxicity for radiotherapy breast cancer treatments

Abstract Purpose/Objective: In order to preserve the patient's cardiac integrity, a good practice that would be simple and quick to implement would be to optimize dose on the left anterior descending coronary artery (LADCA) during dosimetric planning treatment. However, manual delineation of LADCA remains complicated to perform in clinical routine on non-injected CT-scans. To conclude on the feasibility of such a practice, we investigated the automatic delineation robustness of the Mirada software (Mirada Medical, UK) to generate these volumes. Materials/methods: Forty randomly selected CT-scans of breast cancer patients were used. All were delineated by a radiation therapist following international recommendations (Duane et al., 2017). A planning organ at risk volume (PRV) with a 10 mm margin centered on the LADCA was also performed. Among the forty CT-scans, twenty were used to develop the automatic segmentation atlas using “Workflow Box” auto-segmentation algorithm. The remaining twenty other CT-scans were compiled on the software in order to obtain auto-segmented structures. The relevance of the contours was carried out with the Artiview software (Aquilab, Loos-lès-Lille, France) in order to determine both Dice-similarity coefficients (DSC) and Jaccard indexes. Finally, dosimetric planning treatment in volumetric modulated arc therapy conditions was performed using Eclipse treatment planning system (Varian). Results: Because of variability, both in terms of shape and spatial localization, the atlas-based auto-segmentation software did not succeed in LADCA delineation. Then, LADCA contours were not available for DSC and Jaccard indexes assessment. On the other hand, the use of margins on the LADCA made it possible to create such an Atlas. The average DSC is 0.36 ± 0.15 and the Jaccard index reaches 0.23 ± 0.11. The results associated with dosimetry planning treatment are presented in Table 1. The use of a PRV instead of the LADCA contour tends to overestimate the maximum doses by around 30%. However, the mean and median doses are of the same order of magnitude. Conclusion: Small volume structures do not allow the creation of a robust contouring atlas. However, to overcome this limitation, the use of margins to define a PRV has been quite effective. However, it overestimates the maximum doses by about 30%, so the constraints used to limit doses to the coronary arteries during treatment planning will have to be adjusted in this way. Table 1 Dosimetric results from VMAT optimization for breast radiotherapy treatments.Left breastRight breastPRVLADCAPRVLADCAAverage dose (Gy)12.71 ± 6.5413.24 ± 7.485.86 ± 1.345.61 ± 1.39Median Dose (Gy)11.75 ± 6.5713.32 ± 8.335.75 ± 1.325.59 ± 1.27Maximum dose (Gy)33.29 ± 12.8525.28 ± 11.1810.25 ± 1.967.99 ± 2.09 Citation Format: Nicolas Tkatchenko, Pierre Loap, Eliot Nicolas, Youlia Kirova. Evaluation of workflow box mirada software for automatic delineation of left anterior descending coronary artery (LADCA) in order reduce cardiac toxicity for radiotherapy breast cancer treatments [abstract]. In: Proceedings of the 2020 San Antonio Breast Cancer Virtual Symposium; 2020 Dec 8-11; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2021;81(4 Suppl):Abstract nr PS15-16.

Stepwise Diagnosis and Dynamic Change&nbsp;Prediction AI System for COVID-19

Background: As the COVID-19 pandemic continues to spread worldwide, there is still no accurate rapid diagnostic test for COVID-19, or monitoring tool for patient’s clinical course. An artificial-intelligence system (CoviDet) was therefore developed and applied on chest computed tomography (CT), to evaluate its application for rapid diagnosis and potential monitoring of COVID-19 patients. Methods: 1,201,074 CT slices from 2527 patients were grouped into 3 main groups: COVID-19 positive group, non-COVID-19 viral pneumonia group and control group. They were used to train and validate CoviDet novel stepwise diagnostic algorithm to diagnose COVID-19, with or without clinical data. A subset of COVID-19 patients with more than 3 consecutive CT images were selected for training and validation of the auto-segmentation and monitoring algorithm. Findings: CoviDet outperforms radiologists, and can diagnose COVID-19 based on CT alone with sensitivity of 0.93, specificity of 0.95 and AUC of 0.98 (95%CI 0.97-0.99; P<0.001). Its auto-segmentation analyses co-related well with those by radiologists, with a Dice's coefficient of 0.77. Its can generate a predictive curve of patient’s clinical course if serial CT assessments are available. Interpretation: CoviDet is a useful aid for radiologist to diagnose COVID-19 rapidly and accurately based on CT, with or without clinical data. It can produce an objective predictive curve of patient’s clinical course for easy visualisation. In principle, CoviDet could be used on a web-based platform with built auto masking of privacy data, whereby clinicians can upload anonymised CT for COVID-19 analysis without sensitive clinical information, from anywhere in the world. Funding Statement: This study is supported by High-level university construction project of Guangzhou medical university (Grant No. 20182737, 201721007, 201715907, 2017160107); the Guangdong high level hospital construction reaching peak plan. Declaration of Interests: Qiang Du is the Founder and CEO of Beijing XiaoBaiShiJi Network Technical Co. Ltd. All other authors declare no competing interests. Ethics Approval Statement: This study is approved by the ethics committee of the First Affiliated Hospital of Guangzhou Medical University.

A Novel Deep Learning Based Auto Segmentation For Rectum Tumor On MRI Image

Manually contouring gross tumor volume (GTV) is a crucial and time-consuming process in rectum cancer radiotherapy. This study aims to develop a simple deep learning based rectum tumor auto segmentation algorithm on MRI T2 images. MRI scans (3T, T2-weighted) of 93 patients with locally advanced (cT3-4 and/or cN1-2) rectal cancer treated with neoadjuvant chemoradiotherapy followed by surgery were enrolled in this study. A very deep convolutional networks (VGG16) based full convolution network (FCN) was established as training model. The model was trained in two phases to increase efficiency, including tumor recognition and tumor segmentation. An opening (erosion and dilation) process was implemented to make contours smoother. Data were randomly separated into training (90%) and validation (10%) dataset for 10 folder cross-validation. Additionally, 20 patients were double contoured for performance evaluation. Four indices were calculated to evaluate the similarity of automated and manual segmentation, including Hausdorff distance (HD), average surface distance (ASD), Dice index (DSC) and Jaccard index (JSC). The DSC, JSC, HD, ASD (mean±SD) were 0.74±0.14, 0.60±0.16, 20.44±13.35mm and 3.25±1.69mm for validation dataset; And these indices were 0.71±0.13, 0.57±0.15, 14.91±7.62mm and 2.67±1.46mm between two human radiation oncologists. T-test suggested there is no statistically significant difference between automated segmentation and manual segmentation considering DSC (p=0.42), JSC (p=0.35), HD (p=0.079) and ASD (p=0.16). However, significant difference was found for HD (p=0.0027) without opening process. This study showed that a simple deep learning neural network can perform a human comparable segmentation for rectum cancer based on MRI T2 images.

Predicting spirometry readings using cough sound features and regression

Objective: Spirometry is a commonly used method of measuring lung function. It is useful in the definitive diagnosis of diseases such as asthma and chronic obstructive pulmonary disease (COPD). However, spirometry requires cooperative patients, experienced staff, and repeated testing to ensure the consistency of measurements. There is discomfort associated with spirometry and some patients are not able to complete the test. In this paper, we investigate the possibility of using cough sound analysis for the prediction of spirometry measurements. Approach: Our approach is based on the premise that the mechanism of cough generation and the forced expiratory maneuver of spirometry share sufficient similarities enabling this prediction. Using an iPhone, we collected mostly voluntary cough sounds from 322 adults presenting to a respiratory function laboratory for pulmonary function testing. Subjects had the following diagnoses: obstructive, restrictive, or mixed pattern diseases, or were found to have no lung disease along with normal spirometry. The cough sounds were automatically segmented using the algorithm described in Sharan et al ( IEEE Trans. Biomed. Eng.). We then represented cough sounds with various cough sound descriptors and built linear and nonlinear regression models connecting them to spirometry parameters. Augmentation of cough features with subject demographic data is also experimented with. The dataset was divided into 272 training subjects and 50 test subjects for experimentation. Main results: The performance of the auto-segmentation algorithm was evaluated on 49 randomly selected subjects from the overall dataset with a sensitivity and PPV of 84.95% and 98.51%, respectively. Our regression models achieved a root mean square error (and correlation coefficient) for standard spirometry parameters FEV1, FVC, and FEV1/FVC of 0.593L (0.810), 0.725L (0.749), and 0.164 (0.547), respectively, on the test dataset. In addition, we could achieve sensitivity, specificity, and accuracy of 70% or higher by applying the GOLD standard for COPD diagnosis on the estimated spirometry test results. Significance: The experimental results show high positive correlation in predicting FEV1 and FVC and moderate positive correlation in predicting FEV1/FVC. The results show possibility of predicting spirometry results using cough sound analysis.

Development of a robust MRI fiducial system for automated fusion of MR-US abdominal images.

We present the development of a two‐component magnetic resonance (MR) fiducial system, that is, a fiducial marker device combined with an auto‐segmentation algorithm, designed to be paired with existing ultrasound probe tracking and image fusion technology to automatically fuse MR and ultrasound (US) images. The fiducial device consisted of four ~6.4 mL cylindrical wells filled with 1 g/L copper sulfate solution. The algorithm was designed to automatically segment the device in clinical abdominal MR images. The algorithm's detection rate and repeatability were investigated through a phantom study and in human volunteers. The detection rate was 100% in all phantom and human images. The center‐of‐mass of the fiducial device was robustly identified with maximum variations of 2.9 mm in position and 0.9° in angular orientation. In volunteer images, average differences between algorithm‐measured inter‐marker spacings and actual separation distances were 0.53 ± 0.36 mm. “Proof‐of‐concept” automatic MR‐US fusions were conducted with sets of images from both a phantom and volunteer using a commercial prototype system, which was built based on the above findings. Image fusion accuracy was measured to be within 5 mm for breath‐hold scanning. These results demonstrate the capability of this approach to automatically fuse US and MR images acquired across a wide range of clinical abdominal pulse sequences.

An Auto-Segmentation Algorithm for Multi-Label Image Based on Graph Cut

The image segmentation algorithm based on graph cut guarantees a globally optimal solution for energy solution, which is usually with the aid of user’s interactive operation. For the multi-label image segmentation application, the graph cut algorithm has two drawbacks. Firstly, it has a higher computational complexity of segment multi-label images. Secondly, it is prone to be trapped in local minima when solves the energy formulation. For the two drawbacks, this paper presents an auto-segmentation algorithm based on graph cut to segment multi-label images. The number of the labels is obtained via the main colors of the image, then the main colors are employed as pre-specified nodes feature, rather than select seeds with the aid of prior knowledge or initialization operation by the user. The seeds can be selected automatically without complex mathematical formulations to computerize, and it reduces the computational complexity successfully, and avoids falling into local minima effectively. In additional, we use a fast α-expansion move algorithm to optimize the energy function, which can improve the speed of segmentation. Comparing the proposed algorithm with the state-of-the-art segmentation methods, the experimental results show that the proposed algorithm has superior performance.

Accuracy of patient-specific organ dose estimates obtained using an automated image segmentation algorithm.

The overall goal of this work is to develop a rapid, accurate, and automated software tool to estimate patient-specific organ doses from computed tomography (CT) scans using simulations to generate dose maps combined with automated segmentation algorithms. This work quantified the accuracy of organ dose estimates obtained by an automated segmentation algorithm. We hypothesized that the autosegmentation algorithm is sufficiently accurate to provide organ dose estimates, since small errors delineating organ boundaries will have minimal effect when computing mean organ dose. A leave-one-out validation study of the automated algorithm was performed with 20 head-neck CT scans expertly segmented into nine regions. Mean organ doses of the automatically and expertly segmented regions were computed from Monte Carlo-generated dose maps and compared. The automated segmentation algorithm estimated the mean organ dose to be within 10% of the expert segmentation for regions other than the spinal canal, with the median error for each organ region below 2%. In the spinal canal region, the median error was [Formula: see text], with a maximum absolute error of 28% for the single-atlas approach and 11% for the multiatlas approach. The results demonstrate that the automated segmentation algorithm can provide accurate organ dose estimates despite some segmentation errors.

Tissue segmentation of computed tomography images using a Random Forest algorithm: a feasibility study

There is a need for robust, fully automated whole body organ segmentation for diagnostic CT. This study investigates and optimizes a Random Forest algorithm for automated organ segmentation; explores the limitations of a Random Forest algorithm applied to the CT environment; and demonstrates segmentation accuracy in a feasibility study of pediatric and adult patients. To the best of our knowledge, this is the first study to investigate a trainable Weka segmentation (TWS) implementation using Random Forest machine-learning as a means to develop a fully automated tissue segmentation tool developed specifically for pediatric and adult examinations in a diagnostic CT environment.Current innovation in computed tomography (CT) is focused on radiomics, patient-specific radiation dose calculation, and image quality improvement using iterative reconstruction, all of which require specific knowledge of tissue and organ systems within a CT image. The purpose of this study was to develop a fully automated Random Forest classifier algorithm for segmentation of neck–chest–abdomen–pelvis CT examinations based on pediatric and adult CT protocols. Seven materials were classified: background, lung/internal air or gas, fat, muscle, solid organ parenchyma, blood/contrast enhanced fluid, and bone tissue using Matlab and the TWS plugin of FIJI. The following classifier feature filters of TWS were investigated: minimum, maximum, mean, and variance evaluated over a voxel radius of 2n, (n from 0 to 4), along with noise reduction and edge preserving filters: Gaussian, bilateral, Kuwahara, and anisotropic diffusion. The Random Forest algorithm used 200 trees with 2 features randomly selected per node. The optimized auto-segmentation algorithm resulted in 16 image features including features derived from maximum, mean, variance Gaussian and Kuwahara filters. Dice similarity coefficient (DSC) calculations between manually segmented and Random Forest algorithm segmented images from 21 patient image sections, were analyzed. The automated algorithm produced segmentation of seven material classes with a median DSC of 0.86 ± 0.03 for pediatric patient protocols, and 0.85 ± 0.04 for adult patient protocols. Additionally, 100 randomly selected patient examinations were segmented and analyzed, and a mean sensitivity of 0.91 (range: 0.82–0.98), specificity of 0.89 (range: 0.70–0.98), and accuracy of 0.90 (range: 0.76–0.98) were demonstrated. In this study, we demonstrate that this fully automated segmentation tool was able to produce fast and accurate segmentation of the neck and trunk of the body over a wide range of patient habitus and scan parameters.

MO-DE-207B-10: Impact of Morphologic Characteristics On Radiomics Features From Contast-Enhanced CT for Primary Lung Tumors

Purpose: Determine the impact of morphologic characteristics (e.g. necrosis, vascular enhancement, and cavitation) on radiomic features from contrast enhanced CT (CE-CT) in primary lung tumors. Methods: We developed an auto-segmentation algorithm to separate lung tumors on contrast-enhanced CT into cavitation (air), necrosis, tissue, and enhancing vessels using a combination of thresholding and region-growing. An auto-segmentation algorithm was also designed to identify necrosis on FDG-PET scans. Wilcoxon rank-sum tests were used to determine if significant differences existed in radiomics features (histogram-uniformity and Laplacian-of-Gaussian average) from 249 patients, found to prognostic in previous work, based on the presence/absence of morphologic features. Feature values were also compared between the original tumor contours and contours excluding a specific morphologic feature. Comparison of necrosis segmentation on CE-CT versus FDG-PET was performed in 78 patients to assess for agreement using the concordance correlation coefficient (CCC). Results: Tumors with cavitation and enhancing vasculature had lower uniformity values (p = 0.001 and p = 0.03, respectively). Tumors with enhancing vasculature and necrosis had higher Laplacian-of-Gaussian average values (measure of “edges” within the tumor) (p 0.8). This suggests there may be interactions between morphologic characteristics and the radiomic feature value of tumor tissue. Comparison of necrosis volume and percent necrosis volume of tumor were found to have CCC values of 0.85 and 0.76, respectively between CE-CT and FDG-PET segmentation methods. Conclusions: Tumors with enhancing vasculature, necrosis, and cavitation have higher radiomic feature values that are associated with poor prognosis than tumors without these features. Removing these tissue types from quantitative assessment did not drastically impact radiomic feature values. High reproducibility of CE-CT segmented necrosis compared to FDG-PET segmented necrosis provides a reasonable validation of segmentation accuracy on CE-CT.

SU‐C‐207B‐05: Tissue Segmentation of Computed Tomography Images Using a Random Forest Algorithm: A Feasibility Study

Purpose:Develop an automated Random Forest algorithm for tissue segmentation of CT examinations.Methods:Seven materials were classified for segmentation: background, lung/internal gas, fat, muscle, solid organ parenchyma, blood/contrast, and bone using Matlab and the Trainable Weka Segmentation (TWS) plugin of FIJI. The following classifier feature filters of TWS were investigated: minimum, maximum, mean, and variance each evaluated over a pixel radius of 2n, (n = 0–4). Also noise reduction and edge preserving filters, Gaussian, bilateral, Kuwahara, and anisotropic diffusion, were evaluated. The algorithm used 200 trees with 2 features per node. A training data set was established using an anonymized patient's (male, 20 yr, 72 kg) chest‐abdomen‐pelvis CT examination. To establish segmentation ground truth, the training data were manually segmented using Eclipse planning software, and an intra‐observer reproducibility test was conducted. Six additional patient data sets were segmented based on classifier data generated from the training data. Accuracy of segmentation was determined by calculating the Dice similarity coefficient (DSC) between manual and auto segmented images.Results:The optimized autosegmentation algorithm resulted in 16 features calculated using maximum, mean, variance, and Gaussian blur filters with kernel radii of 1, 2, and 4 pixels, in addition to the original CT number, and Kuwahara filter (linear kernel of 19 pixels). Ground truth had a DSC of 0.94 (range: 0.90–0.99) for adult and 0.92 (range: 0.85–0.99) for pediatric data sets across all seven segmentation classes. The automated algorithm produced segmentation with an average DSC of 0.85 ± 0.04 (range: 0.81–1.00) for the adult patients, and 0.86 ± 0.03 (range: 0.80–0.99) for the pediatric patients.Conclusion:The TWS Random Forest auto‐segmentation algorithm was optimized for CT environment, and able to segment seven material classes over a range of body habitus and CT protocol parameters with an average DSC of 0.86 ± 0.04 (range: 0.80–0.99).

Prospective Randomized Double-Blind Study of Atlas-Based Autosegmentation Assisted Radiation Treatment Planning in Head-and-Neck Cancer

We assessed potential time reduction and acceptability of an atlas-based autosegmentation (AS) algorithm compared to manual segmentation (MS) of organs-at-risk (OARs) region of interest (ROI) for radiation therapy (RT) planning.

SU-FF-I-87: DTA-Based Metrics for the Evaluation of Autosegmentation Algorithms in Clinical Radiotherapy Workflow

Purpose: To establish quantitative metrics to assess the performance of autosegmentation programs in the context of clinical radiotherapy workflow. Method and Materials: One of the more time intensive and predominantly un‐automated tasks in radiation therapytreatment planning is the segmentation of the anatomical structures which are relevant for a particular intended radiotherapytreatment. We present performance metrics based on the distance‐to‐agreement (DTA) between manually‐contoured reference sets and the results of a generic autosegmentation platform. Results: To correlate to the time expended to correct erroneous auto‐contouring results we posit the slice‐by‐slice DTA = min { | x r,i − x t,i | } ∀ { x t,i } , where{x r,i} and {x t,i} are the vertices on the reference and test sets, respectively, on the ith axial CT slice. The histogram of DTAs (or simple pass/fail DTA thresholds) will assess the amount of manual correction required to bring an auto‐contoured structure into acceptable agreement with the reference set. Beyond this application, full automation of the contouring process (i.e. no manual correction) would require stricter, 3‐dimensional, and site‐specific DTA metrics, for which we propose generalized quality factors. As a test bed for these metrics, we have compared manually‐generated clinical contours to those created by CMS' Atlas‐Based Autosegmentation system and shown that, for some sites, the autosegmentation is nearly indistinguishable from the variability in our clinic's manual contouring. Conclusion: Depending on the desired use of an autosegmentation package, either slice‐by‐slice or 3D DTA metrics facilitate the assessment of clinical performance. Regardless of intended application, the DTA‐based metrics may be used as (a) a quantitative benchmark of an autosegmentation algorithm against the inherent variability in manual contouring and (b) a potential correlate to efficiency gains should the software be brought into full clinical use.