Learning-based CT Research Articles

CTARR: A fast and robust method for identifying anatomical regions on CT images via atlas registration

Medical image analysis tasks often focus on regions or structures located in a particular location within the patient’s body. Often large parts of the image may not be of interest for the image analysis task. When using deep-learning based approaches, this causes an unnecessary increases the computational burden during inference and raises the chance of errors. In this paper, we introduce CTARR, a novel generic method for CT Anatomical Re- gion Recognition. The method serves as a pre-processing step for any deep learning-based CT image analysis pipeline by automatically identifying the pre-defined anatomical region that is relevant for the follow-up task and removing the rest. It can be used in (i) image segmentation to prevent false positives in anatomically implausible regions and speeding up the inference, (ii) image classification to produce image crops that are consistent in their anatomical context, and (iii) image registration by serving as a fast pre-registration step. Our proposed method is based on atlas registration and provides a fast and robust way to crop any anatomical region encoded as one or multiple bounding box(es) from any unlabeled CT scan of the brain, chest, abdomen and/or pelvis. We demonstrate the utility and robustness of the proposed method in the context of medical image segmentation by evaluating it on six datasets of public segmentation challenges. The foreground voxels in the regions of interest are preserved in the vast majority of cases and tasks (97.45-100%) while taking only fractions of a seconds to compute (0.1-0.21s) on a deep learning worksta- tion and greatly reducing the segmentation runtime (2.0-12.7x). Our code is available at <a href='https://github.com/ThomasBudd/ctarr'>https://github.com/ThomasBudd/ctarr</a>

Deep learning-based CT image for pulmonary nodule classification with intrathoracic fat: A multicenter study

Correctly distinguishing between benign and malignant pulmonary nodules can avoid unnecessary invasive procedures. However, existing models exhibit a deficiency in the study of fat regions. We aim to create a multimodal deep learning nomogram (DLN) using pulmonary nodule and intrathoracic fat (ITF) features to distinguish between benign and malignant nodules and further demonstrate that intrathoracic fat can enhance the performance of lung cancer prediction models. A segmentation methodology was devised for precisely delineating intranodular and perinodular region (IPN). The Swin Transformer was used for feature extraction. The least absolute shrinkage and selection operator (LASSO) regression was employed for feature selection. The nomogram integrating IPN and ITF signatures revealed better diagnostic accuracy than IPN signature alone (net reclassification improvement = 0.134, P<0.05, and integrated discrimination improvement = 0.078, P<0.05). In the internal test cohort, DLN obtained an area under the curve (AUC) of 0.913 (95 % CI: 0.870, 0.952). In the external test cohort, the AUC was 0.906 (95 % CI: 0.867, 0.938). There was no significant difference between the calibrated curve and perfectly calibrated curve in all cohorts (P>0.05). Decision curve analysis showed that the DLN was clinically useful. The DLN demonstrated substantial diagnostic value for distinguishing between benign and malignant pulmonary nodules. The incorporation of the ITF signature has enhanced the performance of DLN.

Image quality of virtual monochromatic and material density iodine images for evaluation of head and neck neoplasms using deep learning-based CT image reconstruction – A retrospective observational study

PurposeTo compare the quality of deep learning image reconstructed (DLIR) virtual monochromatic images (VMI) and material density (MD) iodine images from dual-energy computed tomography (DECT) for the evaluation of head and neck neoplasms with CT scans from a conventional single-energy protocol. MethodA total of 294 head and neck CT scans (98 VMIs operated at 60 keV, 102 MD iodine images, and 94 images from a 120 kVp single-energy CT (SECT) protocol) were retrospectively evaluated. VMIs and MD iodine images were generated using the Gemstone Spectral Imaging (GSI) mode using DLIR and metal artifact reduction (MAR) algorithms. SECT images were generated using adaptive statistical iterative reconstruction (ASIR-V). Images were scored by two independent readers on a 6-point Likert-type scale for overall image quality, vessel contrast, soft tissue contrast, noise texture, noise intensity, artifact reduction, and sharpness. ResultsSubjective overall image quality was rated as superior or excellent in 98 % of DLIR-based MD iodine images and VMIs, but only in 55 % of ASIR-V-based SECT images. For each individual quality criterion, image quality of VMIs and MD iodine images was rated as better than that of SECT images (p < 0.001 in each case). Noise texture and intensity were rated better in MD iodine images than in VMIs. ConclusionDECT using both DLIR and MAR for the generation of VMIs and MD iodine images resulted in higher subjective quality of oncologic head and neck images than ASIR-V-based SECT. Noise reduction and noise texture were best achieved with DLIR-based MD iodine images.

Machine learning-based CT radiomics approach for predicting occult peritoneal metastasis in advanced gastric cancer preoperatively

PurposeTo develop a machine learning-based CT radiomics model to preoperatively diagnose occult peritoneal metastasis (OPM) in advanced gastric cancer (AGC) patients. MethodsA total of 177 AGC patients were retrospectively analyzed. Four region of interests (ROIs) along the largest area of tumor (core ROI) and corresponding tumor mesenteric fat space (peri ROI) were manually delineated on the arterial (A-core and A-peri) and venous phase (V-core and V-peri) of CT images. A total of 1316 radiomics features were extracted from each ROI. Then, ten machine learning classification algorithms were used to develop model. An integrated radiomics nomogram was established to predict OPM individually. ResultsFor the radiomics of tumor mesenteric fat space, the AUCs of A-peri in training set and test set were 0.881 and 0.800. And the AUCs of V-peri were 0.838 and 0.815, respectively. In terms of primary tumor’ s radiomics signature, the AUCs of A-core in training set and test set were 0.862 and 0.691. The AUCs of V-core were 0.831 and 0.620. Integrated radiomics model showed the highest AUC value compare to each single radiomics score in the training (0.943 vs 0.831-0.881) and test set (0.835 vs 0.620-0.815). Radiomics nomogram demonstrated good diagnostic accuracy with a C-index of 0.948. ConclusionsBoth the radiomics of tumor mesenteric fat space and primary tumor were associated with OPM. A CT radiomics nomogram had a relatively good predictive performance for detecting OPM in patients with AGC.

Reconstructing and analyzing the invariances of low-dose CT image denoising networks.

Deep learning-based methods led to significant advancements in many areas of medical imaging, most of which are concerned with the reduction of artifacts caused by motion, scatter, or noise. However, with most neural networks being black boxes, they remain notoriously difficult to interpret, hindering their clinical implementation. In particular, it has been shown that networks exhibit invariances w.r.t. input features, that is, they learn to ignore certain information in the inputdata. To improve the interpretability of deep learning-based low-dose CT image denoisingnetworks. We learn a complete data representation of low-dose input images using a conditional variational autoencoder (cVAE). In this representation, invariances of any given denoising network are then disentangled from the information it is not invariant to using a conditional invertible neural network (cINN). At test time, image-space invariances are generated by applying the inverse of the cINN and subsequent decoding using the cVAE. We propose two methods to analyze sampled invariances and to find those that correspond to alterations of anatomicalstructures. The proposed method is applied to four popular deep learning-based low-dose CT image denoising networks. We find that the networks are not only invariant to noise amplitude and realizations, but also to anatomicalstructures. The proposed method is capable of reconstructing and analyzing invariances of deep learning-based low-dose CT image denoising networks. This is an important step toward interpreting deep learning-based methods for medical imaging, which is essential for their clinical implementation.

Impact of a deep learning-based brain CT interpretation algorithm on clinical decision-making for intracranial hemorrhage in the emergency department

Intracranial hemorrhage is a critical emergency that requires prompt and accurate diagnosis in the emergency department (ED). Deep learning technology can assist in interpreting non-enhanced brain CT scans, but its real-world impact on clinical decision-making is uncertain. This study assessed a deep learning-based intracranial hemorrhage detection algorithm (DLHD) in a simulated clinical environment with ten emergency medical professionals from a tertiary hospital’s ED. The participants reviewed CT scans with clinical information in two steps: without and with DLHD. Diagnostic performance was measured, including sensitivity, specificity, accuracy, and the area under the receiver operating characteristic curve. Consistency in clinical decision-making was evaluated using the kappa statistic. The results demonstrated that DLHD minimally affected experienced participants’ diagnostic performance and decision-making. In contrast, inexperienced participants exhibited significantly increased sensitivity (59.33–72.67%, p < 0.001) and decreased specificity (65.49–53.73%, p < 0.001) with the algorithm. Clinical decision-making consistency was moderate among inexperienced professionals (k = 0.425) and higher among experienced ones (k = 0.738). Inexperienced participants changed their decisions more frequently, mainly due to the algorithm’s false positives. The study highlights the need for thorough evaluation and careful integration of deep learning tools into clinical workflows, especially for less experienced professionals.

Linear diffusion noise boosted deep image prior for unsupervised sparse-view CT reconstruction

Objective. Deep learning has markedly enhanced the performance of sparse-view computed tomography reconstruction. However, the dependence of these methods on supervised training using high-quality paired datasets, and the necessity for retraining under varied physical acquisition conditions, constrain their generalizability across new imaging contexts and settings. Approach. To overcome these limitations, we propose an unsupervised approach grounded in the deep image prior framework. Our approach advances beyond the conventional single noise level input by incorporating multi-level linear diffusion noise, significantly mitigating the risk of overfitting. Furthermore, we embed non-local self-similarity as a deep implicit prior within a self-attention network structure, improving the model’s capability to identify and utilize repetitive patterns throughout the image. Additionally, leveraging imaging physics, gradient backpropagation is performed between the image domain and projection data space to optimize network weights. Main Results. Evaluations with both simulated and clinical cases demonstrate our method’s effective zero-shot adaptability across various projection views, highlighting its robustness and flexibility. Additionally, our approach effectively eliminates noise and streak artifacts while significantly restoring intricate image details. Significance. Our method aims to overcome the limitations in current supervised deep learning-based sparse-view CT reconstruction, offering improved generalizability and adaptability without the need for extensive paired training data.

Development and clinical validation of a deep learning-based knee CT image segmentation method for robotic-assisted total knee arthroplasty.

This study aimed to develop a novel deep convolutional neural network called Dual-path Double Attention Transformer (DDA-Transformer) designed to achieve precise and fast knee joint CT image segmentation and to validate it in robotic-assisted total knee arthroplasty (TKA). The femoral, tibial, patellar, and fibular segmentation performance and speed were evaluated and the accuracy of component sizing, bone resection and alignment of the robotic-assisted TKA system constructed using this deep learning network was clinically validated. Overall, DDA-Transformer outperformed six other networks in terms of the Dice coefficient, intersection over union, average surface distance, and Hausdorff distance. DDA-Transformer exhibited significantly faster segmentation speeds than nnUnet, TransUnet and 3D-Unet (p<0.01). Furthermore, the robotic-assisted TKA system outperforms the manual group in surgical accuracy. DDA-Transformer exhibited significantly improved accuracy and robustness in knee joint segmentation, and this convenient and stable knee joint CT image segmentation network significantly improved the accuracy of the TKA procedure.

Value of CT quantification in progressive fibrosing interstitial lung disease: a deep learning approach.

To evaluate the relationship of changes in the deep learning-based CT quantification of interstitial lung disease (ILD) with changes in forced vital capacity (FVC) and visual assessments of ILD progression, and to investigate their prognostic implications. This study included ILD patients with CT scans at intervals of over 2 years between January 2015 and June 2021. Deep learning-based texture analysis software was used to segment ILD findings on CT images (fibrosis: reticular opacity + honeycombing cysts; total ILD extent: ground-glass opacity + fibrosis). Patients were grouped according to the absolute decline of predicted FVC (< 5%, 5-10%, and ≥ 10%) and ILD progression assessed by thoracic radiologists, and their quantification results were compared among these groups. The associations between quantification results and survival were evaluated using multivariable Cox regression analysis. In total, 468 patients (239 men; 64 ± 9.5 years) were included. Fibrosis and total ILD extents more increased in patients with larger FVC decline (p < .001 in both). Patients with ILD progression had higher fibrosis and total ILD extent increases than those without ILD progression (p < .001 in both). Increases in fibrosis and total ILD extent were significant prognostic factors when adjusted for absolute FVC declines of ≥ 5% (hazard ratio [HR] 1.844, p = .01 for fibrosis; HR 2.484, p < .001 for total ILD extent) and ≥ 10% (HR 2.918, p < .001 for fibrosis; HR 3.125, p < .001 for total ILD extent). Changes in ILD CT quantification correlated with changes in FVC and visual assessment of ILD progression, and they were independent prognostic factors in ILD patients. Quantifying the CT features of interstitial lung disease using deep learning techniques could play a key role in defining and predicting the prognosis of progressive fibrosing interstitial lung disease. • Radiologic findings on high-resolution CT are important in diagnosing progressive fibrosing interstitial lung disease. • Deep learning-based quantification results for fibrosis and total interstitial lung disease extents correlated with the decline in forced vital capacity and visual assessments of interstitial lung disease progression, and emerged as independent prognostic factors. • Deep learning-based interstitial lung disease CT quantification can play a key role in diagnosing and prognosticating progressive fibrosing interstitial lung disease.

Radiomics Analysis of Preprocedural CT Imaging for Outcome Prediction after Transjugular Intrahepatic Portosystemic Shunt Creation.

To evaluate the role of radiomics in preoperative outcome prediction in cirrhotic patients who underwent transjugular intrahepatic portosystemic shunt (TIPS) using "controlled expansion covered stents". This retrospective institutional review board-approved study included cirrhotic patients undergoing TIPS with controlled expansion covered stent placement. From preoperative CT images, the whole liver was segmented into Volumes of Interest (VOIs) at the unenhanced and portal venous phase. Radiomics features were extracted, collected, and analyzed. Subsequently, receiver operating characteristic (ROC) curves were drawn to assess which features could predict patients' outcomes. The endpoints studied were 6-month overall survival (OS), development of hepatic encephalopathy (HE), grade II or higher HE according to West Haven Criteria, and clinical response, defined as the absence of rebleeding or ascites. A radiomic model for outcome prediction was then designed. A total of 76 consecutive cirrhotic patients undergoing TIPS creation were enrolled. The highest performances in terms of the area under the receiver operating characteristic curve (AUROC) were observed for the "clinical response" and "survival at 6 months" outcome with 0.755 and 0.767, at the unenhanced and portal venous phase, respectively. Specifically, on basal scans, accuracy, specificity, and sensitivity were 66.42%, 63.93%, and 73.75%, respectively. At the portal venous phase, an accuracy of 65.34%, a specificity of 62.38%, and a sensitivity of 74.00% were demonstrated. A pre-interventional machine learning-based CT radiomics algorithm could be useful in predicting survival and clinical response after TIPS creation in cirrhotic patients.

Deep learning–based chest CT model to predict treatment response to immune checkpoint inhibitors in non-small cell lung cancer independently and additively to histopathological biomarkers.

8536 Background: Immune checkpoint inhibitors (ICI) are an integral part of the treatment for advanced or metastatic non-small cell lung cancer (NSCLC). Chest CT and pathology data is gathered in routine clinical practice, and each modality provides complementary information. Tumor microenvironment is spatially heterogeneous and temporally dynamic. CT images provide macro-level characteristics of multiple tumors, and tissue-based histopathologic biomarkers analyze a single representative tumor microenvironment in-depth. We developed a deep learning-based chest CT prediction model to predict response to ICIs and validated its performance independently and in combination with histopathologic biomarkers. Methods: NSCLC patients treated with ICI monotherapy were retrospectively enrolled from Samsung Medical Center. A deep learning CT prediction (DL-CT) model was trained on the data of 1876 patients where time to next treatment (TTNT) was used as a proxy. A pre-trained CT nodule detection model was used to extract intermediate features from chest CT taken at a time closest to the initiation of ICI. Histopathologic biomarkers include tumor infiltrating lymphocytes (TIL) and immune phenotype analyzed by an AI analyzer (Lunit SCOPE IO) in H&amp;E whole slide images and PD-L1 tumor proportion score (TPS) interpreted by human pathologists. The performance of DL-CT model was validated in an independent test set of 489 NSCLC patients. Results: In the test set, median follow up was 11.3 months. Patients predicted as a responder by DL-CT model (CT-responder) had prolonged TTNT (median TTNT: 7.0m vs. 2.5m, hazard ratio [HR] 0.588, p &lt;0.001) and overall survival (OS) (median OS: 16.5m vs 7.6m, HR 0.581, p &lt;0.001). The responder group predicted by DL-CT model was associated with better ECOG performance score (p = 0.012) and current smoking (p &lt;0.001), but not with TIL-based immune phenotype or PD-L1 expression (p = 0.201 and p = 0.312, respectively). CT-responder showed favorable outcome across all PD-L1 subgroups (HR for TTNT 0.509, p = 0.005 for TPS 0%, HR 0.616, P = 0.015 for TPS 1-49% and HR 0.585, p &lt;0.001 for TPS ≥50%) and immune phenotypes (HR for TTNT 0.543, p &lt;0.001 for inflamed and HR 0.611, p &lt;0.001 for non-inflamed). Combining DL-CT and histopathologic biomarkers, patients having all three biomarkers positive (CT-responder, inflamed and PDL1 TPS ≥50%) had substantially longer TTNT and OS compared to the other patients (mTTNT 9.4m vs. 3.5m, p&lt;0.001 and mOS 32.0m vs. 10.0m vs. p &lt;0.001). Conclusions: The AI model based on chest CT data was a strong independent biomarker to predict treatment response to ICI, which possibly reflects patients’ clinical characteristics. Combining AI CT model and histopathologic biomarkers showed additive impact on the prediction of the treatment response to ICI.

Computational thinking ability in mathematics learning of exponents in grade IX

In the PISA 2021 framework, Computational Thinking (CT) is described as a detailed mathematical solution to the problem to be solved. However, CT-based learning still needs to be widely applied in Indonesia. This study aims to describe the CT ability of students in grade IX of junior high school based on CT indicators on the material of signed numbers. The data collection techniques in this study were test questions and interviews. Students who obtained high categories with scores above 45.76 were six students with a percentage of 21%, students who received medium categories with scores between 11.94 and 45.76 were 19 students with a rate of 66%, and students who obtained low categories with scores below 11.94 were four people with a percentage of 13%. The results of the study state that as many as 39% of students can decompose the problems given, 17% of students can recognize patterns in the problem, 24% of students can sort out the information in the situation or abstract, and 26% of students can solve problems well according to algorithm indicators.

Accuracy Study on Deep Learning-Based CT Image Analysis for Lung Nodule Detection and Classification

Accuracy Study on Deep Learning-Based CT Image Analysis for Lung Nodule Detection and Classification

Clinical feasibility of deep learning-based synthetic CT images from T2-weighted MR images for cervical cancer patients compared to MRCAT

This work aims to investigate the clinical feasibility of deep learning-based synthetic CT images for cervix cancer, comparing them to MR for calculating attenuation (MRCAT). Patient cohort with 50 pairs of T2-weighted MR and CT images from cervical cancer patients was split into 40 for training and 10 for testing phases. We conducted deformable image registration and Nyul intensity normalization for MR images to maximize the similarity between MR and CT images as a preprocessing step. The processed images were plugged into a deep learning model, generative adversarial network. To prove clinical feasibility, we assessed the accuracy of synthetic CT images in image similarity using structural similarity (SSIM) and mean-absolute-error (MAE) and dosimetry similarity using gamma passing rate (GPR). Dose calculation was performed on the true and synthetic CT images with a commercial Monte Carlo algorithm. Synthetic CT images generated by deep learning outperformed MRCAT images in image similarity by 1.5% in SSIM, and 18.5 HU in MAE. In dosimetry, the DL-based synthetic CT images achieved 98.71% and 96.39% in the GPR at 1% and 1 mm criterion with 10% and 60% cut-off values of the prescription dose, which were 0.9% and 5.1% greater GPRs over MRCAT images.

Comparison of machine learning-based CT fractional flow reserve with cardiac MR perfusion mapping for ischemia diagnosis in stable coronary artery disease.

To compare the diagnostic performance of machine learning (ML)-based computed tomography-derived fractional flow reserve (CT-FFR) and cardiac magnetic resonance (MR) perfusion mapping for functional assessment of coronary stenosis. Between October 2020 and March 2022, consecutive participants with stable coronary artery disease (CAD) were prospectively enrolled and underwent coronary CTA, cardiac MR, and invasive fractional flow reserve (FFR) within 2weeks. Cardiac MR perfusion analysis was quantified by stress myocardial blood flow (MBF) and myocardial perfusion reserve (MPR). Hemodynamically significant stenosis was defined as FFR ≤ 0.8 or > 90% stenosis on invasive coronary angiography (ICA). The diagnostic performance of CT-FFR, MBF, and MPR was compared, using invasive FFR as a reference. The study protocol was completed in 110 participants (mean age, 62years ± 8; 73 men), and hemodynamically significant stenosis was detected in 36 (33%). Among the quantitative perfusion indices, MPR had the largest area under receiver operating characteristic curve (AUC) (0.90) for identifying hemodynamically significant stenosis, which is in comparison with ML-based CT-FFR on the vessel level (AUC 0.89, p = 0.71), with comparable sensitivity (89% vs 79%, p = 0.20), specificity (87% vs 84%, p = 0.48), and accuracy (88% vs 83%, p = 0.24). However, MPR outperformed ML-based CT-FFR on the patient level (AUC 0.96 vs 0.86, p = 0.03), with improved specificity (95% vs 82%, p = 0.01) and accuracy (95% vs 81%, p < 0.01). ML-based CT-FFR and quantitative cardiac MR showed comparable diagnostic performance in detecting vessel-specific hemodynamically significant stenosis, whereas quantitative perfusion mapping had a favorable performance in per-patient analysis. ML-based CT-FFR and MPR derived from cardiac MR performed well in diagnosing vessel-specific hemodynamically significant stenosis, both of which showed no statistical discrepancy with each other. • Both machine learning (ML)-based computed tomography-derived fractional flow reserve (CT-FFR) and quantitative perfusion cardiac MR performed well in the detection of hemodynamically significant stenosis. • Compared with stress myocardial blood flow (MBF) from quantitative perfusion cardiac MR, myocardial perfusion reserve (MPR) provided higher diagnostic performance for detecting hemodynamically significant coronary artery stenosis. • ML-based CT-FFR and MPR from quantitative cardiac MR perfusion yielded similar diagnostic performance in assessing vessel-specific hemodynamically significant stenosis, whereas MPR had a favorable performance in per-patient analysis.

Machine learning-based CT texture analysis in the differentiation of testicular masses.

To evaluate the ability of texture features for distinguishing between benign and malignant testicular masses, and furthermore, for identifying primary testicular lymphoma in malignant tumors and identifying seminoma in testicular germ cell tumors, respectively. We retrospectively collected 77 patients with an abdominal and pelvic enhanced computed tomography (CT) examination and a histopathologically confirmed testicular mass from a single center. The ROI of each mass was split into two parts by the largest cross-sectional slice and deemed to be two samples. After all processing steps, three-dimensional texture features were extracted from unenhanced and contrast-enhanced CT images. Excellent reproducibility of texture features was defined as intra-class correlation coefficient ≥0.8 (ICC ≥0.8). All the groups were balanced via the synthetic minority over-sampling technique (SMOTE) method. Dimension reduction was based on pearson correlation coefficient (PCC). Before model building, minimum-redundancy maximum-relevance (mRMR) selection and recursive feature elimination (RFE) were used for further feature selection. At last, three ML classifiers with the highest cross validation with 5-fold were selected: autoencoder (AE), support vector machine(SVM), linear discriminant analysis (LAD). Logistics regression (LR) and LR-LASSO were also constructed to compare with the ML classifiers. 985 texture features with ICC ≥0.8 were extracted for further feature selection process. With the highest AUC of 0.946 (P <0.01), logistics regression was proved to be the best model for the identification of benign or malignant testicular masses. Besides, LR also had the best performance in identifying primary testicular lymphoma in malignant testicular tumors and in identifying seminoma in testicular germ cell tumors, with the AUC of 0.982 (P <0.01) and 0.928 (P <0.01), respectively. Until now, this is the first study that applied CT texture analysis (CTTA) to assess the heterogeneity of testicular tumors. LR model based on CTTA might be a promising non-invasive tool for the diagnosis and differentiation of testicular masses. The accurate diagnosis of testicular masses would assist urologists in correct preoperative and perioperative decision making.

MIS-Net: A deep learning-based multi-class segmentation model for CT images.

The accuracy of traditional CT image segmentation algorithms is hindered by issues such as low contrast and high noise in the images. While numerous scholars have introduced deep learning-based CT image segmentation algorithms, they still face challenges, particularly in achieving high edge accuracy and addressing pixel classification errors. To tackle these issues, this study proposes the MIS-Net (Medical Images Segment Net) model, a deep learning-based approach. The MIS-Net model incorporates multi-scale atrous convolution into the encoding and decoding structure with symmetry, enabling the comprehensive extraction of multi-scale features from CT images. This enhancement aims to improve the accuracy of lung and liver edge segmentation. In the evaluation using the COVID-19 CT Lung and Infection Segmentation dataset, the left and right lung segmentation results demonstrate that MIS-Net achieves a Dice Similarity Coefficient (DSC) of 97.61. Similarly, in the Liver Tumor Segmentation Challenge 2017 public dataset, the DSC of MIS-Net reaches 98.78.

Evaluating the Hounsfield unit assignment and dose differences between CT-based standard and deep learning-based synthetic CT images for MRI-only radiation therapy of the head and neck.

Magnetic resonance image only (MRI-only) simulation for head and neck (H&N) radiotherapy (RT) could allow for single-image modality planning with excellent soft tissue contrast. In the MRI-only simulation workflow, synthetic computed tomography (sCT) is generated from MRI to provide electron density information for dose calculation. Bone/air regions produce little MRI signal which could lead to electron density misclassification in sCT. Establishing the dosimetric impact of this error could inform quality assurance (QA) procedures using MRI-only RT planning or compensatory methods for accurate dosimetric calculation. The aim of this study was to investigate if Hounsfield unit (HU) voxel misassignments from sCT images result in dosimetric errors in clinical treatment plans. Fourteen H&N cancer patients undergoing same-day CT and 3T MRI simulation were retrospectively identified. MRI was deformed to the CT using multimodal deformable image registration. sCTs were generated from T1w DIXON MRIs using a commercially available deep learning-based generator (MRIplanner, Spectronic Medical AB, Helsingborg, Sweden). Tissue voxel assignment was quantified by creating a CT-derived HU threshold contour. CT/sCT HU differences for anatomical/target contours and tissue classification regions including air (<250 HU), adipose tissue (-250 HU to -51 HU), soft tissue (-50 HU to 199 HU), spongy (200 HU to 499 HU) and cortical bone(>500 HU) were quantified. t-test was used to determine if sCT/CT HU differences were significant. The frequency of structures that had a HU difference>80 HU (the CT window-width setting for intra-cranial structures) was computed to establish structure classification accuracy. Clinical intensity modulated radiation therapy (IMRT) treatment plans created on CT were retrospectively recalculated on sCT images and compared using the gamma metric. The mean ratio of sCT HUs relative to CT for air, adipose tissue, soft tissue, spongy and cortical bone were 1.7±0.3, 1.1±0.1, 1.0±0.1, 0.9±0.1 and 0.8±0.1 (value of 1 indicates perfect agreement). T-tests (significance set at t=0.05) identified differences in HU values for air, spongy and cortical bone in sCT images compared to CT. The structures with sCT/CT HU differences>80 HU of note were the left and right (L/R) cochlea and mandible (>79% of the tested cohort), the oral cavity (for 57% of the tested cohort), the epiglottis (for 43% of the tested cohort) and the L/R TM joints (occurring>29% of the cohort). In the case of the cochlea and TM joints, these structures contain dense bone/air interfaces. In the case of the oral cavity and mandible, these structures suffer the additional challenge of being positionally altered in CT versus MRI simulation (due to a non-MR safe immobilizing bite block requiring absence of bite block in MR). Finally, the epiglottis HU assignment suffers from its small size and unstable positionality. Plans recalculated on sCT yielded global/local gamma pass rates of 95.5%±2% (3mm, 3%) and 92.7%±2.1% (2mm, 2%). The largest mean differences in D95, Dmean , D50 dose volume histogram (DVH) metrics for organ-at-risk (OAR) and planning tumor volumes (PTVs) were 2.3%±3.0% and 0.7%±1.9% respectively. In this cohort, HU differences of CT and sCT were observed but did not translate into a reduction in gamma pass rates or differences in average PTV/OAR dose metrics greater than 3%. For sites such as the H&N where there are many tissue interfaces we did not observe large scale dose deviations but further studies using larger retrospective cohorts are merited to establish the variation in sCT dosimetric accuracy which could help to inform QA limits on clinical sCT usage.

Quantifying the Dosimetric Impact of Tissue Hounsfield Unit Assignment in Deep Learning-based Synthetic CT Images For MRI-Only Radiation Therapy of The Head and Neck

MRI-only simulation for head and neck (HN) radiotherapy (RT) could allow for single-image modality planning with excellent soft tissue contrast. In the MRI-only simulation workflow, synthetic CTs (sCTs) are generated from MRI to provide electron density information for dose calculation. Bone/air regions produce little MRI signal which could lead to electron density misclassification in sCTs. Establishing the dosimetric impact of this error could inform quality assurance (QA) procedures using MRI-only RT planning. In this study we quantify differences in Hounsfield Unit (HU) values between paired CT/ sCTs of HN cancer patients and investigate the dosimetric impact on clinical treatment plans. Fourteen patients with head and neck cancer undergoing same-day CT and 3T MRI simulation were retrospectively identified. MRIs were deformed to the CT using multimodal deformable image registration. SCTs were generated from T1w DIXON MRI using a commercially available deep learning-based generator (MRIplanner, Spectronics). Tissue voxel assignment was quantified by creating a CT-derived HU threshold contour. CT/sCT HU differences for anatomical/target contours and tissue classification regions including air (<-250HU), adipose tissue (-250HU to -51HU), soft tissue (-50HU to 199HU), spongy (200HU to 499HU) and cortical bone (> 500HU) were quantified. T-test was used to determine if sCT/CT HU differences were significant. The frequency of structures that had a HU difference >70HU (the CT window-width setting for intra-cranial structures) was computed to establish structure classification accuracy. Clinical IMRT treatment plans created on CTs were retrospectively recalculated on sCT images using compared using the gamma metric. The mean ratio of sCT HUs relative to CT for air, adipose tissue, soft tissue, spongy and cortical bone were 1.7±0.3, 1.1±0.1, 1.0±0.1, 0.9±0.1 and 0.8±0.1 (value of 1 indicates perfect agreement). T-tests (significance set at t = 0.05) identified differences in HU values for air, spongy and cortical bone in sCT images compared to CT. The structures with sCT/CT HU differences > 70HU were the L/R cochlea and mandible, occurring in >78% of the tested cohort. These structures contain dense bone/air interfaces. Plans recalculated on sCTs yielded global/local gamma pass rates of 98.7%±1.2% (3mm,3%) and 95.5%±2.5% (1mm,1%). Mean differences in D95, D50, D10 and D2 dose volume histogram (DVH) metrics for organ-at-risk (OAR) and planning tumor volumes (PTV) were 0.8%±1.5% and 2.1% ± 1.2% respectively. In this cohort, HU differences in sCTs were observed but did not translate into a reduction in gamma pass rates and OAR/PTV DVH metrics. The acquisition of additional training data such as ultrashort echo time MRI could improve bone/air contrast and reduce bone/air sCT misclassifications. Further studies will establish the variation in sCT dosimetric accuracy using a larger retrospective cohort to inform QA limits on clinical sCT usage.

Clinical Feasibility of Deep Learning-Based CT during Treatment CBCT Tumor Registration-Segmentation in Thoracic Radiotherapy (RT)

Accurate tumor segmentation on weekly cone-beam computed tomography (CBCT) images is critical for image-guided and adaptive radiation therapy (ART). In thoracic RT, low image contrast, imaging artifact, and geometry and image modality differences from the planning CT (pCT) typically limits accurate tumor segmentation and registration. Here, we explored the clinical feasibility of using 3D recurrent registration-segmentation deep learning (DL) that combines patient-specific anatomic and shape context from higher contrast pCT and planning contours (PACs) for tumor segmentation on during treatment CBCTs. We included the pCT and CBCTs from six patients with locally advanced non-small cell lung cancer (LA-NSCLC) who had underwent RT. Cases were selected with a primary GTV contoured and labeled separately from the nodal GTV. Using rigidly aligned pCT and CBCT as inputs, DL auto-segmented the GTV on week 1 and 6 CBCTs, and these auto-segmented contours were manually inspected by a radiation oncologist that edited the GTV according to clinical standard quality. The Dice similarity coefficient (DSC), Hausdorff distance (HD95), mean surface distance (MSD), surface DSC (sDSC) and added path length (APL) were used to quantitively compare the DL and the edited GTV. The primary GTV was in the right lung in five cases, and left lung in one case. Manual adjustments were typically made at the interface of GTV and lung parenchyma with partial inclusion of adjacent vessels. Hypodensities within the GTV were sometimes not segmented in all axial slices resulting in discontinuous components. The quantitative comparison between the edited and DL-generated GTV is shown in Table 1. For week 1, the average DSC and HD95 were 0.87 and 6.94 mm, respectively. The performance for week 6 was slightly lower than week 1, with a DSC of 0.85 and HD95 of 7.22 mm. The agreement with the generated DL GTV and the edited GTV was high in week 1 and decreased somewhat later during the treatment course possibly due to a higher impact of geometric changes in tumor and adjacent structures. The proposed DL algorithm showed reasonable performance throughout the treatment, supporting its potential for use into clinical routine for LA-NSCLC.