Dual-region radiomics signature: Integrating primary tumor and lymph node computed tomography features improves survival prediction in esophageal squamous cell cancer

Dual-Region Computed Tomography Radiomics-Based Machine Learning Predicts Subcarinal Lymph Node Metastasis in Patients with Non-small Cell Lung Cancer

BACKGROUND

Noninvasively and accurately predicting subcarinal lymph node metastasis (SLNM) for patients with non-small cell lung cancer (NSCLC) remains challenging. This study was designed to develop and validate a tumor and subcarinal lymph nodes (tumor-SLNs) dual-region computed tomography (CT) radiomics model for predicting SLNM in NSCLC.

METHODS

This retrospective study included NSCLC patients who underwent lung resection and SLNs dissection between January 2017 and December 2020. The radiomic features of the tumor and SLNs were extracted from preoperative CT, respectively. Ninety machine learning (ML) models were developed based on tumor region, SLNs region, and tumor-SLNs dual-region. The model performance was assessed by the area under the curve (AUC) and validated internally by fivefold cross-validation.

RESULTS

In total, 202 patients were included in this study. ML models based on dual-region radiomics showed good performance for SLNM prediction, with a median AUC of 0.794 (range, 0.686-0.880), which was superior to those of models based on tumor region (median AUC, 0.746; range, 0.630-0.811) and SLNs region (median AUC, 0.700; range, 0.610-0.842). The ML model, which is developed by using the naive Bayes algorithm and dual-region features, had the highest AUC of 0.880 (range of cross-validation, 0.825-0.937) among all ML models. The optimal logistic regression model was inferior to the optimal ML model for predicting SLNM, with an AUC of 0.727.

CONCLUSIONS

The CT radiomics showed the potential for accurately predicting SLNM in NSCLC patients. The ML model with dual-region radiomic features has better performance than the logistic regression or single-region models.

Advances in the Imaging of Esophageal and Gastroesophageal Junction Malignancies

Accurate imaging is key for the diagnosis and treatment of esophageal and gastroesophageal junction cancers . Current imaging modalities, such as computed tomography (CT) and 18F-FDG (2-deoxy-2-[18F]fluoro-D-glucose) positron emission tomography (PET)/CT, have limitations in accurately staging these cancers. MRI shows promise for T staging and residual disease assessment. Novel PET tracers, like FAPI, FLT, and hypoxia markers, offer potential improvements in diagnostic accuracy. 18F-FDG PET/MRI combines metabolic and anatomic information, enhancing disease evaluation. Radiomics and artificial intelligence hold promise for early detection, treatment planning, and response assessment. Theranostic nanoparticles and personalized medicine approaches offer new avenues for cancer therapy.

Intratumoral and peritumoral CT-based radiomics for predicting the microsatellite instability in gastric cancer

PURPOSE

To investigate the value of intratumoral and peritumoral radiomics based on contrast-enhanced computer tomography (CECT) to preoperatively predict microsatellite instability (MSI) status in gastric cancer (GC) patients.

METHODS

A total of 189 GC patients, including 63 patients with MSI-high (MSI-H) and 126 patients with MSI-low/stable (MSI-L/S), were randomly divided into the training cohort and validation cohort. Intratumoral and 5-mm peritumoral regions' radiomics features were extracted from CECT images. The features were standardized by Z-score, and the Inter- and intraclass correlation coefficient, univariate logistic regression analysis, and least absolute shrinkage and selection operator (LASSO) were applied to select the optimal radiomics features. Radiomics scores (Rad-score) based on intratumoral regions, peritumoral regions, and intratumoral + 5-mm peritumoral regions were calculated by weighting the linear combination of the selected features with their respective coefficients to construct the intratumoral model, peritumoral model, and intratumoral + peritumoral model. Logistic regression was used to establish a combined model by combining clinical characteristics, CT semantic features, and Rad-score of intratumoral and peritumoral regions.

RESULTS

Eleven radiomics features were selected to establish a radiomics intratumoral + peritumoral model. CT-measured tumor length and tumor location were independent risk factors for MSI status. The established combined model obtained the highest area under the receiver operating characteristic (ROC) curve (AUC) of 0.830 (95% CI, 0.727-0.906) in the validation cohort. The calibration curve and decision curve demonstrated its good model fitness and clinical application value.

CONCLUSION

The combined model based on intratumoral and peritumoral CECT radiomics features and clinical factors can predict the MSI status of GS with moderate accuracy before surgery, which helps formulate personalized treatment strategies.

A combined nomogram based on radiomics and hematology to predict the pathological complete response of neoadjuvant immunochemotherapy in esophageal squamous cell carcinoma

BACKGROUND

To predict pathological complete response (pCR) in patients receiving neoadjuvant immunochemotherapy (nICT) for esophageal squamous cell carcinoma (ESCC), we explored the factors that influence pCR after nICT and established a combined nomogram model.

METHODS

We retrospectively included 164 ESCC patients treated with nICT. The radiomics signature and hematology model were constructed utilizing least absolute shrinkage and selection operator (LASSO) regression, and the radiomics score (radScore) and hematology score (hemScore) were determined for each patient. Using the radScore, hemScore, and independent influencing factors obtained through univariate and multivariate analyses, a combined nomogram was established. The consistency and prediction ability of the nomogram were assessed utilizing calibration curve and the area under the receiver operating factor curve (AUC), and the clinical benefits were assessed utilizing decision curve analysis (DCA).

RESULTS

We constructed three predictive models.The AUC values of the radiomics signature and hematology model reached 0.874 (95% CI: 0.819-0.928) and 0.772 (95% CI: 0.699-0.845), respectively. Tumor length, cN stage, the radScore, and the hemScore were found to be independent factors influencing pCR according to univariate and multivariate analyses (P < 0.05). A combined nomogram was constructed from these factors, and AUC reached 0.934 (95% CI: 0.896-0.972). DCA demonstrated that the clinical benefits brought by the nomogram for patients across an extensive range were greater than those of other individual models.

CONCLUSIONS

By combining CT radiomics, hematological factors, and clinicopathological characteristics before treatment, we developed a nomogram model that effectively predicted whether ESCC patients would achieve pCR after nICT, thus identifying patients who are sensitive to nICT and assisting in clinical treatment decision-making.

Preoperative Prediction of Perineural Invasion in Oesophageal Squamous Cell Carcinoma Based on CT Radiomics Nomogram: A Multicenter Study

RATIONALE AND OBJECTIVES

To investigate the value of computed tomography (CT) radiomics nomogram in the preoperative prediction of perineural invasion (PNI) in oesophageal squamous cell carcinoma (ESCC) through a multicenter study.

MATERIALS AND METHODS

We retrospectively collected postoperative pathological data of 360 ESCC patients with definite PNI status (131 PNI-positive and 229 PNI-negative) from two centres. Radiomic features were extracted from the arterial-phase CT images, and the least absolute shrinkage and selection operator and logistic regression algorithm were used to screen valuable features for identifying the PNI status and calculating the radiomics score (Rad-score). A radiomics nomogram was established by integrating the Rad-score and clinical risk factors. A receiver operating characteristic curve was used to evaluate model performance, and decision curve analysis was used to evaluate the predictive performance of the radiomics nomogram in the training, internal validation, and external validation sets.

RESULTS

Twenty radiomics features were extracted from a full-volume tumour region of interest to construct the model, and the radiomics nomogram combined with radiomics features and clinical risk factors was superior to the clinical and radiomics models in predicting the PNI status of ESCC patients. The area under the curve values of the radiomics nomogram in the training, internal validation, and external validation sets were 0.856 (0.794-0.918), 0.832 (0.742-0.922), and 0.803 (0.709-0.898), respectively.

CONCLUSION

The radiomics nomogram based on CT has excellent predictive ability; it can non-invasively predict the preoperative PNI status of ESCC patients and provide a basis for preoperative decision-making.

Preoperative CT radiomics of esophageal squamous cell carcinoma and lymph node to predict nodal disease with a high diagnostic capability

PURPOSE

To develop CT radiomics models of resectable esophageal squamous cell carcinoma (ESCC) and lymph node (LN) to preoperatively identify LN+.

MATERIALS AND METHODS

299 consecutive patients with ESCC were enrolled in the study, 140 of whom were LN+ and 159 were LN-. Of the 299 patients, 249 (from the same hospital) were randomly divided into a training cohort (n = 174) and a test cohort (n = 75). The remaining 50 patients, from a second hospital, were assigned to an external validation cohort. In the training cohort, preoperative contrast-enhanced CT radiomics features of ESCC and LN were extracted, then integrated with clinical features to develop three models: ESCC, LN and combined. The performance of these models was assessed using area under receiver operating characteristic curve (AUC), and F-1 score, which were validated in both the test cohort and external validation cohort.

RESULTS

An ESCC model was developed for the training cohort utilizing the 8 tumor radiomics features, and an LN model was constructed using 9 nodal radiomics features. A combined model was constructed using both ESCC and LN extracted features, in addition to cT stage and LN+ distribution. This combined model had the highest predictive ability among the three models in the training cohort (AUC = 0.948, F1-score = 0.878). The predictive ability was validated in both the test and external validation cohorts (AUC = 0.885 and 0.867, F1-score = 0.816 and 0.773, respectively).

CONCLUSION

To preoperatively determine LN+, the combined model is superior to models of ESCC and LN alone.

Radiomics Nomogram with Added Nodal Features Improves Treatment Response Prediction in Locally Advanced Esophageal Squamous Cell Carcinoma: A Multicenter Study

OBJECTIVE

We aimed to develop and validate a radiomics nomogram and determine the value of radiomic features from lymph nodes (LNs) for predicting pathological complete response (pCR) to neoadjuvant chemoradiotherapy (NCRT) in patients with locally advanced esophageal squamous cell carcinoma (ESCC).

METHODS

In this multicenter retrospective study, eligible participants who had undergone NCRT followed by radical esophagectomy were consecutively recruited. Three radiomics models (modelT, modelLN, and modelTLN) based on tumor and LN features, alone and combined, were developed in the training cohort. The radiomics nomogram was developed by incorporating the prediction value of the radiomics model and clinicoradiological risk factors using multivariate logistic regression, and was evaluated using the receiver operating characteristic curve, validated in two external validation cohorts.

RESULTS

Between October 2011 and December 2018, 116 patients were included in the training cohort. Between June 2015 and October 2020, 51 and 27 patients from two independent hospitals were included in validation cohorts 1 and 2, respectively. The radiomics modelTLN performed better than the radiomics modelT for predicting pCR. The radiomics nomogram incorporating the predictive value of the radiomics modelTLN and heterogeneous after NCRT outperformed the clinicoradiological model, with an area under the curve (95% confidence interval) of 0.833 (0.765-0.894) versus 0.764 (0.686-0.833) [p = 0.088, DeLong test], 0.824 (0.718-0.909) versus 0.692 (0.554-0.809) [p = 0.012], and 0.902 (0.794-0.984) versus 0.696 (0.526-0.857) [p = 0.024] in all three cohorts.

CONCLUSIONS

Radiomic features from LNs could provide additional value for predicting pCR in ESCC patients, and the radiomics nomogram provided an accurate prediction of pCR, which might aid treatment decision.

A radiomics nomogram for predicting postoperative recurrence in esophageal squamous cell carcinoma

Purpose

To establish and validate a radiomics nomogram for predicting recurrence of esophageal squamous cell carcinoma (ESCC) after esophagectomy with curative intent.

Materials and methods

The medical records of 155 patients who underwent surgical treatment for pathologically confirmed ESCC were collected. Patients were randomly divided into a training group (n=109) and a validation group (n=46) in a 7:3 ratio. ​Tumor regions are accurately segmented in computed tomography images of enrolled patients. Radiomic features were then extracted from the segmented tumors. We selected the features by Max-relevance and min-redundancy (mRMR) and least absolute shrinkage and selection operator (LASSO) methods. A radiomics signature was then built by logistic regression analysis. To improve predictive performance, a radiomics nomogram that incorporated the radiomics signature and independent clinical predictors was built. Model performance was evaluated by receiver operating characteristic (ROC) curve, calibration curve, and decision curve analyses (DCA).

Results

We selected the five most relevant radiomics features to construct the radiomics signature. The radiomics model had general discrimination ability with an area under the ROC curve (AUC) of 0.79 in the training set that was verified by an AUC of 0.76 in the validation set. The radiomics nomogram consisted of the radiomics signature, and N stage showed excellent predictive performance in the training and validation sets with AUCs of 0.85 and 0.83, respectively. Furthermore, calibration curves and the DCA analysis demonstrated good fit and clinical utility of the radiomics nomogram.

Conclusion

We successfully established and validated a prediction model that combined radiomics features and N stage, which can be used to predict four-year recurrence risk in patients with ESCC who undergo surgery.

Predicting response to CCRT for esophageal squamous carcinoma by a radiomics-clinical SHAP model

BACKGROUND

Radical concurrent chemoradiotherapy (CCRT) is frequently used as the first-line treatment for patients with locally advanced esophageal cancer. Unfortunately, some patients respond poorly. To predict response to radical concurrent chemoradiotherapy in pre-treatment patients with esophageal squamous carcinoma (ESCC), and compare the predicting efficacies of radiomics features of primary tumor with or without regional lymph nodes, we developed a radiomics-clinical model based on the positioning CT images. Finally, SHapley Additive exPlanation (SHAP) was used to explain the models.

METHODS

This retrospective study enrolled 105 patients with medically inoperable and/or unresectable ESCC who underwent radical concurrent chemoradiotherapy (CCRT) between October 2018 and May 2023. Patients were classified into responder and non-responder groups with RECIST standards. The 11 recently admitted patients were chosen as the validation set, previously admitted patients were randomly split into the training set (n = 70) and the testing set (n = 24). Primary tumor site (GTV), the primary tumor and the uninvolved lymph nodes at risk of microscopic disease (CTV) were identified as Regions of Interests (ROIs). 1762 radiomics features from GTV and CTV were respectively extracted and then filtered by statistical differential analysis and Least Absolute Shrinkage and Selection Operator (LASSO). The filtered radiomics features combined with 13 clinical features were further filtered with Mutual Information (MI) algorithm. Based on the filtered features, we developed five models (Clinical Model, GTV Model, GTV-Clinical Model, CTV Model, and CTV-Clinical Model) using the random forest algorithm and evaluated for their accuracy, precision, recall, F1-Score and AUC. Finally, SHAP algorithm was adopted for model interpretation to achieve transparency and utilizability.

RESULTS

The GTV-Clinical model achieves an AUC of 0.82 with a 95% confidence interval (CI) of 0.76-0.99 on testing set and an AUC of 0.97 with a 95% confidence interval (CI) of 0.84-1.0 on validation set, which are significantly higher than those of other models in predicting ESCC response to CCRT. The SHAP force map provides an integrated view of the impact of each feature on individual patients, while the SHAP summary plots indicate that radiomics features have a greater influence on model prediction than clinical factors in our model.

CONCLUSION

GTV-Clinical model based on texture features and the maximum diameter of lesion (MDL) may assist clinicians in pre-treatment predicting ESCC response to CCRT.

CT radiomics features of meso-esophageal fat in predicting overall survival of patients with locally advanced esophageal squamous cell carcinoma treated by definitive chemoradiotherapy

OBJECTIVE

To investigate the value of CT radiomics features of meso-esophageal fat in the overall survival (OS) prediction of patients with locally advanced esophageal squamous cell carcinoma (ESCC).

METHODS

A total of 166 patients with locally advanced ESCC in two medical centers were retrospectively analyzed. The volume of interest (VOI) of meso-esophageal fat and tumor were manually delineated on enhanced chest CT using ITK-SNAP. Radiomics features were extracted from the VOIs by Pyradiomics and then selected using the t-test, the Cox regression analysis, and the least absolute shrinkage and selection operator. The radiomics scores of meso-esophageal fat and tumors for OS were constructed by a linear combination of the selected radiomic features. The performance of both models was evaluated and compared by the C-index. Time-dependent receiver operating characteristic (ROC) analysis was employed to analyze the prognostic value of the meso-esophageal fat-based model. A combined model for risk evaluation was constructed based on multivariate analysis.

RESULTS

The CT radiomic model of meso-esophageal fat showed valuable performance for survival analysis, with C-indexes of 0.688, 0.708, and 0.660 in the training, internal, and external validation cohorts, respectively. The 1-year, 2-year, and 3-year ROC curves showed AUCs of 0.640-0.793 in the cohorts. The model performed equivalently compared to the tumor-based radiomic model and performed better compared to the CT features-based model. Multivariate analysis showed that meso-rad-score was the only factor associated with OS.

CONCLUSIONS

A baseline CT radiomic model based on the meso-esophagus provide valuable prognostic information for ESCC patients treated with dCRT.

Development and Validation of a New Staging System for Esophageal Squamous Cell Carcinoma Patients Based on Combined Pathological TNM, Radiomics, and Proteomics

OBJECTIVE

This study aimed to construct a new staging system for patients with esophageal squamous cell carcinoma (ESCC) based on combined pathological TNM (pTNM) stage, radiomics, and proteomics.

METHODS

This study collected patients with radiomics and pTNM stage (Cohort 1, n = 786), among whom 103 patients also had proteomic data (Cohort 2, n = 103). The Cox regression model with the least absolute shrinkage and selection operator, and the Cox proportional hazards model were used to construct a nomogram and predictive models. Concordance index (C-index) and the integrated area under the time-dependent receiver operating characteristic (ROC) curve (IAUC) were used to evaluate the predictive models. The corresponding staging systems were further assessed using Kaplan-Meier survival curves.

RESULTS

For Cohort 1, the RadpTNM4c staging systems, constructed based on combined pTNM stage and radiomic features, outperformed the pTNM4c stage in both the training dataset 1 (Train1; IAUC 0.711 vs. 0.706, p < 0.001) and the validation dataset 1 (Valid1; IAUC 0.695 vs. 0.659, p < 0.001; C-index 0.703 vs. 0.674, p = 0.029). For Cohort 2, the ProtRadpTNM2c staging system, constructed based on combined pTNM stage, radiomics, and proteomics, outperformed the pTNM2c stage in both the Train2 (IAUC 0.777 vs. 0.610, p < 0.001; C-index 0.898 vs. 0.608, p < 0.001) and Valid2 (IAUC 0.746 vs. 0.608, p < 0.001; C-index 0.889 vs. 0.641, p = 0.009) datasets.

CONCLUSIONS

The ProtRadpTNM2c staging system, based on combined pTNM stage, radiomic, and proteomic features, improves the predictive performance of the classical pTNM staging system.

CIABNet: Category imbalance attention block network for the classification of multi-differentiated types of esophageal cancer

BACKGROUND

Esophageal cancer has become one of the important cancers that seriously threaten human life and health, and its incidence and mortality rate are still among the top malignant tumors. Histopathological image analysis is the gold standard for diagnosing different differentiation types of esophageal cancer.

PURPOSE

The grading accuracy and interpretability of the auxiliary diagnostic model for esophageal cancer are seriously affected by small interclass differences, imbalanced data distribution, and poor model interpretability. Therefore, we focused on developing the category imbalance attention block network (CIABNet) model to try to solve the previous problems.

METHODS

First, the quantitative metrics and model visualization results are integrated to transfer knowledge from the source domain images to better identify the regions of interest (ROI) in the target domain of esophageal cancer. Second, in order to pay attention to the subtle interclass differences, we propose the concatenate fusion attention block, which can focus on the contextual local feature relationships and the changes of channel attention weights among different regions simultaneously. Third, we proposed a category imbalance attention module, which treats each esophageal cancer differentiation class fairly based on aggregating different intensity information at multiple scales and explores more representative regional features for each class, which effectively mitigates the negative impact of category imbalance. Finally, we use feature map visualization to focus on interpreting whether the ROIs are the same or similar between the model and pathologists, thus better improving the interpretability of the model.

RESULTS

The experimental results show that the CIABNet model outperforms other state-of-the-art models, which achieves the most advanced results in classifying the differentiation types of esophageal cancer with an average classification accuracy of 92.24%, an average precision of 93.52%, an average recall of 90.31%, an average F1 value of 91.73%, and an average AUC value of 97.43%. In addition, the CIABNet model has essentially similar or identical to the ROI of pathologists in identifying histopathological images of esophageal cancer.

CONCLUSIONS

Our experimental results prove that our proposed computer-aided diagnostic algorithm shows great potential in histopathological images of multi-differentiated types of esophageal cancer.

A CT-Based Radiomics Nomogram Model for Differentiating Primary Malignant Melanoma of the Esophagus from Esophageal Squamous Cell Carcinoma

Objective

The diagnosis of primary malignant melanoma of the esophagus (PMME) before treatment is essential for clinical decision-making. However, PMME may be misdiagnosed as esophageal squamous cell carcinoma (ESCC) sometimes. This research is aimed at devising a radiomics nomogram model of CT for distinguishing PMME from ESCC.

Methods

In this retrospective analysis, 122 individuals with proven pathologically PMME (n = 28) and ESCC (n = 94) were registered from our hospital. PyRadiomics was applied to derive radiomics features from plain and enhanced CT images after resampling image into an isotropic resolution of 0.625 × 0.625 × 0.625 mm³. The diagnostic efficiency of the model was evaluated by an independent validation group.

Results

For the purpose of differentiation between PMME and ESCC, a radiomics model was constructed using 5 radiomics features obtained from nonenhanced CT and 4 radiomics features derived from enhanced CT. A radiomics model including multiple radiomics features showed excellent discrimination efficiency with AUCs of 0.975 and 0.906 in the primary and validation cohorts, respectively. Then, a radiomics nomogram model was developed. The decision curve analysis has shown remarkable performance of this nomogram model for distinguishing PMME from ESCC.

Conclusions

The proposed radiomics nomogram model based on CT could be used for distinguishing PMME from ESCC. Moreover, this model also contributed to helping clinicians determine an appropriate treatment strategy for esophageal neoplasms.

The application of radiomics in esophageal cancer: Predicting the response after neoadjuvant therapy

Esophageal cancer (EC) is one of the fatal malignant neoplasms worldwide. Neoadjuvant therapy (NAT) combined with surgery has become the standard treatment for locally advanced EC. However, the treatment efficacy for patients with EC who received NAT varies from patient to patient. Currently, the evaluation of efficacy after NAT for EC lacks accurate and uniform criteria. Radiomics is a multi-parameter quantitative approach for developing medical imaging in the era of precision medicine and has provided a novel view of medical images. As a non-invasive image analysis method, radiomics is an inevitable trend in NAT efficacy prediction and prognosis classification of EC by analyzing the high-throughput imaging features of lesions extracted from medical images. In this literature review, we discuss the definition and workflow of radiomics, the advances in efficacy prediction after NAT, and the current application of radiomics for predicting efficacy after NAT.

Role of radiomics in the diagnosis and treatment of gastrointestinal cancer

Gastrointestinal cancer (GIC) has high morbidity and mortality as one of the main causes of cancer death. Preoperative risk stratification is critical to guide patient management, but traditional imaging studies have difficulty predicting its biological behavior. The emerging field of radiomics allows the conversion of potential pathophysiological information in existing medical images that cannot be visually recognized into high-dimensional quantitative image features. Tumor lesion characterization, therapeutic response evaluation, and survival prediction can be achieved by analyzing the relationships between these features and clinical and genetic data. In recent years, the clinical application of radiomics to GIC has increased dramatically. In this editorial, we describe the latest progress in the application of radiomics to GIC and discuss the value of its potential clinical applications, as well as its limitations and future directions.

Can 18F-FDG PET/CT Radiomics Features Predict Clinical Outcomes in Patients with Locally Advanced Esophageal Squamous Cell Carcinoma?

Simple Summary

PET/CT is an important staging modality in the baseline assessment of locally advanced esophageal squamous cell carcinoma. Accurate staging and response prediction in these patients is essential for management. The aim of this retrospective study was to assess the usefulness of ¹⁸F-FDG PET/CT radiomics features in predicting outcomes such as tumor and nodal categories, PET-based response to induction chemotherapy, progression-free survival, and overall survival. In a final cohort of 74 patients, we found that the developed radiomics models can predict these clinical and prognostic outcomes with reasonable accuracy, similar or better than those derived from conventional imaging. Future studies with a larger cohort would be helpful in establishing the significance of these models.

Abstract

This study aimed to assess the usefulness of radiomics features of ¹⁸F-FDG PET/CT in patients with locally advanced esophageal cancers (ESCC) in predicting outcomes such as clinical tumor (cT) and nodal (cN) categories, PET response to induction chemotherapy (PET response), progression-free survival (PFS), and overall survival (OS). Pretreatment PET/CT images from patients who underwent concurrent chemoradiotherapy from July 2002 to February 2017 were segmented, and data were split into training and test sets. Model development was performed on the training datasets and a maximum of five features were selected. Final diagnostic accuracies were determined using the test dataset. A total of 86 PET/CTs (58 men and 28 women, mean age 65 years) were segmented. Due to small lesion size, 12 patients were excluded. The diagnostic accuracies as derived from the CT, PET, and combined PET/CT test datasets were as follows: cT category—70.4%, 70.4%, and 81.5%, respectively; cN category—69.0%, 86.2%, and 86.2%, respectively; PET response—60.0%, 66.7%, and 70.0%, respectively; PFS—60.7%, 75.0%, and 75.0%, respectively; and OS—51.7%, 55.2%, and 62.1%, respectively. A radiomics assessment of locally advanced ESCC has the potential to predict various clinical outcomes. External validation of these models would be further helpful.

Can 18F-FDG PET/CT Radiomics Features Predict Clinical Outcomes in Patients with Locally Advanced Esophageal Squamous Cell Carcinoma?

This study aimed to assess the usefulness of radiomics features of ¹⁸F-FDG PET/CT in patients with locally advanced esophageal cancers (ESCC) in predicting outcomes such as clinical tumor (cT) and nodal (cN) categories, PET response to induction chemotherapy (PET response), progression-free survival (PFS), and overall survival (OS). Pretreatment PET/CT images from patients who underwent concurrent chemoradiotherapy from July 2002 to February 2017 were segmented, and data were split into training and test sets. Model development was performed on the training datasets and a maximum of five features were selected. Final diagnostic accuracies were determined using the test dataset. A total of 86 PET/CTs (58 men and 28 women, mean age 65 years) were segmented. Due to small lesion size, 12 patients were excluded. The diagnostic accuracies as derived from the CT, PET, and combined PET/CT test datasets were as follows: cT category-70.4%, 70.4%, and 81.5%, respectively; cN category-69.0%, 86.2%, and 86.2%, respectively; PET response-60.0%, 66.7%, and 70.0%, respectively; PFS-60.7%, 75.0%, and 75.0%, respectively; and OS-51.7%, 55.2%, and 62.1%, respectively. A radiomics assessment of locally advanced ESCC has the potential to predict various clinical outcomes. External validation of these models would be further helpful.

Testing the Ability of Convolutional Neural Networks to Learn Radiomic Features

BACKGROUND AND OBJECTIVE

Radiomics and deep learning have emerged as two distinct approaches to medical image analysis. However, their relative expressive power remains largely unknown. Theoretically, hand-crafted radiomic features represent a mere subset of features that neural networks can approximate, thus making deep learning a more powerful approach. On the other hand, automated learning of hand-crafted features may require a prohibitively large number of training samples. Here we directly test the ability of convolutional neural networks (CNNs) to learn and predict the intensity, shape, and texture properties of tumors as defined by standardized radiomic features.

METHODS

Conventional 2D and 3D CNN architectures with an increasing number of convolutional layers were trained to predict the values of 16 standardized radiomic features from real and synthetic PET images of tumors, and tested. In addition, several ImageNet-pretrained advanced networks were tested. A total of 4000 images were used for training, 500 for validation, and 500 for testing.

RESULTS

Features quantifying size and intensity were predicted with high accuracy, while shape irregularity and heterogeneity features had very high prediction errors and generalized poorly. For example, mean normalized prediction error of tumor diameter with a 5-layer CNN was 4.23 ± 0.25, while the error for tumor sphericity was 15.64 ± 0.93. We additionally found that learning shape features required an order of magnitude more samples compared to intensity and size features.

CONCLUSIONS

Our findings imply that CNNs trained to perform various image-based clinical tasks may generally under-utilize the shape and texture information that is more easily captured by radiomics. We speculate that to improve the CNN performance, shape and texture features can be computed explicitly and added as auxiliary variables to the networks, or supplied as synthetic inputs.