Comparative Multicentric Evaluation of Inter-Observer Variability in Manual and Automatic Segmentation of Neuroblastic Tumors in Magnetic Resonance Images

Impact of signal-to-noise ratio and contrast definition on the sensitivity assessment and benchmarking of fluorescence molecular imaging systems

Significance

Standardization of fluorescence molecular imaging (FMI) is critical for ensuring quality control in guiding surgical procedures. To accurately evaluate system performance, two metrics, the signal-to-noise ratio (SNR) and contrast, are widely employed. However, there is currently no consensus on how these metrics can be computed.

Aim

We aim to examine the impact of SNR and contrast definitions on the performance assessment of FMI systems.

Approach

We quantified the SNR and contrast of six near-infrared FMI systems by imaging a multi-parametric phantom. Based on approaches commonly used in the literature, we quantified seven SNRs and four contrast values considering different background regions and/or formulas. Then, we calculated benchmarking (BM) scores and respective rank values for each system.

Results

We show that the performance assessment of an FMI system changes depending on the background locations and the applied quantification method. For a single system, the different metrics can vary up to ∼ 35    dB (SNR), ∼ 8.65    a . u . (contrast), and ∼ 0.67    a . u . (BM score).

Conclusions

The definition of precise guidelines for FMI performance assessment is imperative to ensure successful clinical translation of the technology. Such guidelines can also enable quality control for the already clinically approved indocyanine green-based fluorescence image-guided surgery.

Physics-informed deep generative learning for quantitative assessment of the retina

Disruption of retinal vasculature is linked to various diseases, including diabetic retinopathy and macular degeneration, leading to vision loss. We present here a novel algorithmic approach that generates highly realistic digital models of human retinal blood vessels, based on established biophysical principles, including fully-connected arterial and venous trees with a single inlet and outlet. This approach, using physics-informed generative adversarial networks (PI-GAN), enables the segmentation and reconstruction of blood vessel networks with no human input and which out-performs human labelling. Segmentation of DRIVE and STARE retina photograph datasets provided near state-of-the-art vessel segmentation, with training on only a small (n = 100) simulated dataset. Our findings highlight the potential of PI-GAN for accurate retinal vasculature characterization, with implications for improving early disease detection, monitoring disease progression, and improving patient care.

Reproducibility Analysis of Radiomic Features on T2-weighted MR Images after Processing and Segmentation Alterations in Neuroblastoma Tumors

Purpose To evaluate the reproducibility of radiomics features extracted from T2-weighted MR images in patients with neuroblastoma. Materials and Methods A retrospective study included 419 patients (mean age, 29 months ± 34 [SD]; 220 male, 199 female) with neuroblastic tumors diagnosed between 2002 and 2023, within the scope of the PRedictive In-silico Multiscale Analytics to support cancer personalized diaGnosis and prognosis, Empowered by imaging biomarkers (ie, PRIMAGE) project, involving 746 T2/T2*-weighted MRI sequences at diagnosis and/or after initial chemotherapy. Images underwent processing steps (denoising, inhomogeneity bias field correction, normalization, and resampling). Tumors were automatically segmented, and 107 shape, first-order, and second-order radiomics features were extracted, considered as the reference standard. Subsequently, the previous image processing settings were modified, and volumetric masks were applied. New radiomics features were extracted and compared with the reference standard. Reproducibility was assessed using the concordance correlation coefficient (CCC); intrasubject repeatability was measured using the coefficient of variation (CoV). Results When normalization was omitted, only 5% of the radiomics features demonstrated high reproducibility. Statistical analysis revealed significant changes in the normalization and resampling processes (P < .001). Inhomogeneities removal had the least impact on radiomics (83% of parameters remained stable). Shape features remained stable after mask modifications, with a CCC greater than 0.90. Mask modifications were the most favorable changes for achieving high CCC values, with a radiomics features stability of 70%. Only 7% of second-order radiomics features showed an excellent CoV of less than 0.10. Conclusion Modifications in the T2-weighted MRI preparation process in patients with neuroblastoma resulted in changes in radiomics features, with normalization identified as the most influential factor for reproducibility. Inhomogeneities removal had the least impact on radiomics features. Keywords: Pediatrics, MR Imaging, Oncology, Radiomics, Reproducibility, Repeatability, Neuroblastic Tumors Supplemental material is available for this article. © RSNA, 2024 See also the commentary by Safdar and Galaria in this issue.

Imaging biomarkers and radiomics in pediatric oncology: a view from the PRIMAGE (PRedictive In silico Multiscale Analytics to support cancer personalized diaGnosis and prognosis, Empowered by imaging biomarkers) project

This review paper presents the practical development of imaging biomarkers in the scope of the PRIMAGE (PRedictive In silico Multiscale Analytics to support cancer personalized diaGnosis and prognosis, Empowered by imaging biomarkers) project, as a noninvasive and reliable way to improve the diagnosis and prognosis in pediatric oncology. The PRIMAGE project is a European multi-center research initiative that focuses on developing medical imaging-derived artificial intelligence (AI) solutions designed to enhance overall management and decision-making for two types of pediatric cancer: neuroblastoma and diffuse intrinsic pontine glioma. To allow this, the PRIMAGE project has created an open-cloud platform that combines imaging, clinical, and molecular data together with AI models developed from this data, creating a comprehensive decision support environment for clinicians managing patients with these two cancers. In order to achieve this, a standardized data processing and analysis workflow was implemented to generate robust and reliable predictions for different clinical endpoints. Magnetic resonance (MR) image harmonization and registration was performed as part of the workflow. Subsequently, an automated tool for the detection and segmentation of tumors was trained and internally validated. The Dice similarity coefficient obtained for the independent validation dataset was 0.997, indicating compatibility with the manual segmentation variability. Following this, radiomics and deep features were extracted and correlated with clinical endpoints. Finally, reproducible and relevant imaging quantitative features were integrated with clinical and molecular data to enrich both the predictive models and a set of visual analytics tools, making the PRIMAGE platform a complete clinical decision aid system. In order to ensure the advancement of research in this field and to foster engagement with the wider research community, the PRIMAGE data repository and platform are currently being integrated into the European Federation for Cancer Images (EUCAIM), which is the largest European cancer imaging research infrastructure created to date.

A narrative review of radiomics and deep learning advances in neuroblastoma: updates and challenges

Neuroblastoma is an extremely heterogeneous tumor that commonly occurs in children. The diagnosis and treatment of this tumor pose considerable challenges due to its varied clinical presentations and intricate genetic aberrations. Presently, various imaging modalities, including computed tomography, magnetic resonance imaging, and positron emission tomography, are utilized to assess neuroblastoma. Nevertheless, these conventional imaging modalities have limitations in providing quantitative information for accurate diagnosis and prognosis. Radiomics, an emerging technique, can extract intricate medical imaging information that is imperceptible to the human eye and transform it into quantitative data. In conjunction with deep learning algorithms, radiomics holds great promise in complementing existing imaging modalities. The aim of this review is to showcase the potential of radiomics and deep learning advancements to enhance the diagnostic capabilities of current imaging modalities for neuroblastoma.

Lack of incremental value of three-dimensional measurement in assessing invasiveness for lung cancer

OBJECTIVES

The aim of this study was to evaluate the performance of consolidation-to-tumour ratio (CTR) and the radiomic models in two- and three-dimensional modalities for assessing radiological invasiveness in early-stage lung adenocarcinoma.

METHODS

A retrospective analysis was conducted on patients with early-stage lung adenocarcinoma from Guangdong Provincial People's Hospital and Shenzhen People's Hospital. Manual delineation of pulmonary nodules along the boundary was performed on cross-sectional images to extract radiomic features. Clinicopathological characteristics and radiomic signatures were identified in both cohorts. CTR and radiomic score for every patient were calculated. The performance of CTR and radiomic models were tested and validated in the respective cohorts.

RESULTS

A total of 818 patients from Guangdong Provincial People's Hospital were included in the primary cohort, while 474 patients from Shenzhen People's Hospital constituted an independent validation cohort. Both CTR and radiomic score were identified as independent factors for predicting pathological invasiveness. CTR in two- and three-dimensional modalities exhibited comparable results with areas under the receiver operating characteristic curves and were demonstrated in the validation cohort (area under the curve: 0.807 vs 0.826, P = 0.059) Furthermore, both CTR in two- and three-dimensional modalities was able to stratify patients with significant relapse-free survival (P < 0.000 vs P < 0.000) and overall survival (P = 0.003 vs P = 0.001). The radiomic models in two- and three-dimensional modalities demonstrated favourable discrimination and calibration in independent cohorts (P = 0.189).

CONCLUSIONS

Three-dimensional measurement provides no additional clinical benefit compared to two-dimensional.

Neural network for autonomous segmentation and volumetric assessment of clot and edema in acute and subacute intracerebral hemorrhages

INTRODUCTION

Minimally-invasive surgical techniques for intracerebral hemorrhage (ICH) evacuation use imaging to guide the suction, lysing and/or drainage from the hemorrhage site via various designs. A previous international surgical study has shown that reduction of hematoma volume below 15 ml is indicative of improved long term patient outcomes. The study noted a need for tools to periodically visualize remaining clot during intervention to increase the likelihood of evacuating sufficient clot volumes without endangering rebleeds. Robust segmentation of MRI could guide surgeons and radiologists regarding remaining regions and approaches for prudent evacuation. We thus propose a Convolutional Neural Network (CNN) to identify and autonomously segment clot and peripheral edema in MR images of the brain and generate an estimate of the remaining clot volume.

MATERIALS AND METHODS

We used a retrospective, locally-acquired dataset of ICH patient scans taken on 3 T MRI scanners. Three sets of ground truth manual segmentations were independently generated by two imaging scientists and one radiology fellow. Evaluation of clot age was determined based on relative contrast of hemorrhage components and reviewed by a neurosurgeon. Model accuracy was determined by pixel-wise Dice coefficient (DC) calculations between each ground truth manual segmentation and the machine-derived autonomous segmentations.

RESULTS

The model produced autonomous segmentations of clot core with an average DC of 0.75 ± 0.21 relative to manual segmentations of the same scans. For edema, it produced segmentations with an average DC of 0.68 ± 0.16 relative to manual. From these pixel-wise segmentations, clot volume can be calculated. Model-produced segmentations underestimated clot volumes by an average of 17% relative to ground-truth.

CONCLUSION

The machine learning models were able to identify and segment volumes of ICH components swiftly and accurately.

A multiscale orchestrated computational framework to reveal emergent phenomena in neuroblastoma

Neuroblastoma is a complex and aggressive type of cancer that affects children. Current treatments involve a combination of surgery, chemotherapy, radiotherapy, and stem cell transplantation. However, treatment outcomes vary due to the heterogeneous nature of the disease. Computational models have been used to analyse data, simulate biological processes, and predict disease progression and treatment outcomes. While continuum cancer models capture the overall behaviour of tumours, and agent-based models represent the complex behaviour of individual cells, multiscale models represent interactions at different organisational levels, providing a more comprehensive understanding of the system. In 2018, the PRIMAGE consortium was formed to build a cloud-based decision support system for neuroblastoma, including a multi-scale model for patient-specific simulations of disease progression. In this work we have developed this multi-scale model that includes data such as patient's tumour geometry, cellularity, vascularization, genetics and type of chemotherapy treatment, and integrated it into an online platform that runs the simulations on a high-performance computation cluster using Onedata and Kubernetes technologies. This infrastructure will allow clinicians to optimise treatment regimens and reduce the number of costly and time-consuming clinical trials. This manuscript outlines the challenging framework's model architecture, data workflow, hypothesis, and resources employed in its development.

Prior knowledge based deep learning auto-segmentation in magnetic resonance imaging-guided radiotherapy of prostate cancer

Background and purpose

Automation is desirable for organ segmentation in radiotherapy. This study compared deep learning methods for auto-segmentation of organs-at-risk (OARs) and clinical target volume (CTV) in prostate cancer patients undergoing fractionated magnetic resonance (MR)-guided adaptive radiation therapy. Models predicting dense displacement fields (DDFMs) between planning and fraction images were compared to patient-specific (PSM) and baseline (BM) segmentation models.

Materials and methods

A dataset of 92 patients with planning and fraction MR images (MRIs) from two institutions were used. DDFMs were trained to predict dense displacement fields (DDFs) between the planning and fraction images, which were subsequently used to propagate the planning contours of the bladder, rectum, and CTV to the daily MRI. The training was performed either with true planning-fraction image pairs or with planning images and their counterparts deformed by known DDFs. The BMs were trained on 53 planning images, while to generate PSMs, the BMs were fine-tuned using the planning image of a given single patient. The evaluation included Dice similarity coefficient (DSC), the average (HDavg) and the 95th percentile (HD95) Hausdorff distance (HD).

Results

The DDFMs with DSCs for bladder/rectum of 0.76/0.76 performed worse than PSMs (0.91/0.90) and BMs (0.89/0.88). The same trend was observed for HDs. For CTV, DDFM and PSM performed similarly yielding DSCs of 0.87 and 0.84, respectively.

Conclusions

DDFMs were found suitable for CTV delineation after rigid alignment. However, for OARs they were outperformed by PSMs, as they predicted only limited deformations even in the presence of substantial anatomical changes.

Automatic Segmentation with Deep Learning in Radiotherapy

This review provides a formal overview of current automatic segmentation studies that use deep learning in radiotherapy. It covers 807 published papers and includes multiple cancer sites, image types (CT/MRI/PET), and segmentation methods. We collect key statistics about the papers to uncover commonalities, trends, and methods, and identify areas where more research might be needed. Moreover, we analyzed the corpus by posing explicit questions aimed at providing high-quality and actionable insights, including: "What should researchers think about when starting a segmentation study?", "How can research practices in medical image segmentation be improved?", "What is missing from the current corpus?", and more. This allowed us to provide practical guidelines on how to conduct a good segmentation study in today's competitive environment that will be useful for future research within the field, regardless of the specific radiotherapeutic subfield. To aid in our analysis, we used the large language model ChatGPT to condense information.

Mathematical Model of Clonal Evolution Proposes a Personalised Multi-Modal Therapy for High-Risk Neuroblastoma

Neuroblastoma is the most common extra-cranial solid tumour in children. Despite multi-modal therapy, over half of the high-risk patients will succumb. One contributing factor is the one-size-fits-all nature of multi-modal therapy. For example, during the first step (induction chemotherapy), the standard regimen (rapid COJEC) administers fixed doses of chemotherapeutic agents in eight two-week cycles. Perhaps because of differences in resistance, this standard regimen results in highly heterogeneous outcomes in different tumours. In this study, we formulated a mathematical model comprising ordinary differential equations. The equations describe the clonal evolution within a neuroblastoma tumour being treated with vincristine and cyclophosphamide, which are used in the rapid COJEC regimen, including genetically conferred and phenotypic drug resistance. The equations also describe the agents’ pharmacokinetics. We devised an optimisation algorithm to find the best chemotherapy schedules for tumours with different pre-treatment clonal compositions. The optimised chemotherapy schedules exploit the cytotoxic difference between the two drugs and intra-tumoural clonal competition to shrink the tumours as much as possible during induction chemotherapy and before surgical removal. They indicate that induction chemotherapy can be improved by finding and using personalised schedules. More broadly, we propose that the overall multi-modal therapy can be enhanced by employing targeted therapies against the mutations and oncogenic pathways enriched and activated by the chemotherapeutic agents. To translate the proposed personalised multi-modal therapy into clinical use, patient-specific model calibration and treatment optimisation are necessary. This entails a decision support system informed by emerging medical technologies such as multi-region sequencing and liquid biopsies. The results and tools presented in this paper could be the foundation of this decision support system.

Whole-tumor radiomics analysis of T2-weighted imaging in differentiating neuroblastoma from ganglioneuroblastoma/ganglioneuroma in children: an exploratory study

PURPOSE

To examine the potential of whole-tumor radiomics analysis of T2-weighted imaging (T2WI) in differentiating neuroblastoma (NB) from ganglioneuroblastoma/ganglioneuroma (GNB/GN) in children.

MATERIALS AND METHODS

This study included 102 children with peripheral neuroblastic tumors, comprising 47 NB patients and 55 GNB/GN patients, which were randomly divided into a training group (n = 72) and a test group (n = 30). Radiomics features were extracted from T2WI images, and feature dimensionality reduction was applied. Linear discriminant analysis was used to construct radiomics models, and one-standard error role combined with leave-one-out cross-validation was used to choose the optimal radiomics model with the least predictive error. Subsequently, the patient age at initial diagnosis and the selected radiomics features were incorporated to construct a combined model. The receiver operator characteristic (ROC) curve, decision curve analysis (DCA) and clinical impact curve (CIC) were applied to evaluate the diagnostic performance and clinical utility of the models.

RESULTS

Fifteen radiomics features were eventually chosen to construct the optimal radiomics model. The area under the curve (AUC) of the radiomics model in the training group and test group was 0.940 [95% confidence interval (CI) 0.886, 0.995] and 0.799 (95%CI 0.632, 0.966), respectively. The combined model, which incorporated patient age and radiomics features, achieved an AUC of 0.963 (95%CI 0.925, 1.000) in the training group and 0.871 (95%CI 0.744, 0.997) in the test group. DCA and CIC demonstrated that the radiomics model and combined model could provide benefits at various thresholds, with the combined model being superior to the radiomics model.

CONCLUSION

Radiomics features derived from T2WI, in combination with the age of the patient at initial diagnosis, may offer a quantitative method for distinguishing NB from GNB/GN, thus aiding in the pathological differentiation of peripheral neuroblastic tumors in children.

Independent Validation of a Deep Learning nnU-Net Tool for Neuroblastoma Detection and Segmentation in MR Images

OBJECTIVES

To externally validate and assess the accuracy of a previously trained fully automatic nnU-Net CNN algorithm to identify and segment primary neuroblastoma tumors in MR images in a large children cohort.

METHODS

An international multicenter, multivendor imaging repository of patients with neuroblastic tumors was used to validate the performance of a trained Machine Learning (ML) tool to identify and delineate primary neuroblastoma tumors. The dataset was heterogeneous and completely independent from the one used to train and tune the model, consisting of 300 children with neuroblastic tumors having 535 MR T2-weighted sequences (486 sequences at diagnosis and 49 after finalization of the first phase of chemotherapy). The automatic segmentation algorithm was based on a nnU-Net architecture developed within the PRIMAGE project. For comparison, the segmentation masks were manually edited by an expert radiologist, and the time for the manual editing was recorded. Different overlaps and spatial metrics were calculated to compare both masks.

RESULTS

The median Dice Similarity Coefficient (DSC) was high 0.997; 0.944-1.000 (median; Q1-Q3). In 18 MR sequences (6%), the net was not able neither to identify nor segment the tumor. No differences were found regarding the MR magnetic field, type of T2 sequence, or tumor location. No significant differences in the performance of the net were found in patients with an MR performed after chemotherapy. The time for visual inspection of the generated masks was 7.9 ± 7.5 (mean ± Standard Deviation (SD)) seconds. Those cases where manual editing was needed (136 masks) required 124 ± 120 s.

CONCLUSIONS

The automatic CNN was able to locate and segment the primary tumor on the T2-weighted images in 94% of cases. There was an extremely high agreement between the automatic tool and the manually edited masks. This is the first study to validate an automatic segmentation model for neuroblastic tumor identification and segmentation with body MR images. The semi-automatic approach with minor manual editing of the deep learning segmentation increases the radiologist's confidence in the solution with a minor workload for the radiologist.