Deep radiomics-based fusion model for prediction of bevacizumab treatment response and outcome in patients with colorectal cancer liver metastases: a multicentre cohort study

Multimodal deep learning approaches for precision oncology: a comprehensive review

The burgeoning accumulation of large-scale biomedical data in oncology, alongside significant strides in deep learning (DL) technologies, has established multimodal DL (MDL) as a cornerstone of precision oncology. This review provides an overview of MDL applications in this field, based on an extensive literature survey. In total, 651 articles published before September 2024 are included. We first outline publicly available multimodal datasets that support cancer research. Then, we discuss key DL training methods, data representation techniques, and fusion strategies for integrating multimodal data. The review also examines MDL applications in tumor segmentation, detection, diagnosis, prognosis, treatment selection, and therapy response monitoring. Finally, we critically assess the limitations of current approaches and propose directions for future research. By synthesizing current progress and identifying challenges, this review aims to guide future efforts in leveraging MDL to advance precision oncology.

Estimating Efficacy of Conversion Therapy on Patients with Initially Unresectable Colorectal Cancer Liver Metastases by using MRI: Development of a Predictive Score

RATIONALE AND OBJECTIVES

The conversion success rate (CSR) has crucial implication for clinical outcomes of initially unresectable colorectal liver metastases (CRLM) following conversion therapy. This study aimed to develop a simple predictive scoring model for identifying CSR according to baseline magnetic resonance imaging (MRI) features, and confirm its performance and prognostic significance in a validation cohort.

METHODS

A total of 155 consecutive patients with initially unresectable CRLM were retrospectively reviewed in the study. A simple MRI-based predictive scoring model for identifying CSR was developed in the development cohort (n = 104) by using multivariable logistic regression analyzes. The diagnostic performance was evaluated for the predictive score. Thereafter, patients in the validation cohort (n = 51) were stratified into groups with predicted high CSR or low CSR according to the score. The progression-free survival (PFS) and overall survival (OS) were compared between two groups using the log-rank test.

RESULTS

The predictive score of CSR, named mrNISE, incorporated the number of CRLM ≥ 10, the largest size ≥ 50 mm, poorly defined tumor-liver interface, and peritumoral enhancement. The AUC of the mrNISE score was 0.845 for the development cohort and 0.776 for the validation cohort. According to the score, patients with predicted high CSR had better PFS and OS than those with low CSR in both development and validation cohorts.

CONCLUSION

The predictive score demonstrated great performance for identifying CSR of initially unresectable CRLM. Stratifying patients by the score, personalized treatment goals can be formulated before conversion therapy to improve clinical prognosis and reduce adverse events caused by ineffective treatment.

Predicting prognosis for epithelial ovarian cancer patients receiving bevacizumab treatment with CT-based deep learning

Epithelial ovarian cancer (EOC) presents considerable difficulties in prognostication and treatment strategy development. Bevacizumab, an anti-angiogenic medication, has demonstrated potential in enhancing progression-free survival (PFS) in EOC patients. Nevertheless, the identification of individuals at elevated risk of disease progression following treatment remains a challenging task. This study was to develop and validate a deep learning (DL) model using retrospectively collected computed tomography (CT) plain scans of inoperable and recurrent EOC patients receiving bevacizumab treatment diagnosed between January 2013 and January 2024. A total of 525 patients from three different institutions were retrospectively included in the study and divided into training set (N = 400), internal test set (N = 97) and external test set (N = 28). The model's performance was evaluated using Harrell's C-index. Patients were categorized into high-risk and low-risk group based on a predetermined cutoff in the training set. Additionally, a multimodal model was evaluated, incorporating the risk score generated by the DL model and the pretreatment level of carbohydrate antigen 125 as input variables. The Net Reclassification Improvement (NRI) metric quantified the reclassification performance of our optimal model in comparison to the International Federation of Gynecology and Obstetrics (FIGO) staging model. The results indicated that DL model achieved a PFS predictive C-index of 0.73 in the internal test set and a C-index of 0.61 in the external test set, along with hazard ratios of 34.24 in the training set (95% CI: 21.7, 54.1; P < 0.001) and 8.16 in the internal test set (95% CI: 2.5, 26.8; P < 0.001). The multimodal model demonstrated a C-index of 0.76 in the internal test set and a C-index of 0.64 in the external test set. Comparative analysis against FIGO staging revealed an NRI of 0.06 (P < 0.001) for the multimodal model. The model presents opportunities for prognostic assessment, treatment strategizing, and ongoing patient monitoring.

A CT-based radiomics tumor quality and quantity model to predict early recurrence after radical surgery for colorectal liver metastases

PURPOSE

This study aimed to develop a tumor radiomics quality and quantity model (RQQM) based on preoperative enhanced CT to predict early recurrence after radical surgery for colorectal liver metastases (CRLM).

METHODS

A retrospective analysis was conducted on 282 cases from 3 centers. Clinical risk factors were examined using univariate and multivariate logistic regression (LR) to construct the clinical model. Radiomics features were extracted using the least absolute shrinkage and selection operator (LASSO) for dimensionality reduction. The LR learning algorithm was employed to construct the radiomics model, RQQM (radiomics-TBS), combined model (radiomics-clinical), clinical risk score (CRS) model and tumor burden score (TBS) model. Inter-model comparisons were made using area under the curve (AUC), decision curve analysis (DCA) and calibration curve. Log-rank tests assessed differences in disease-free survival (DFS) and overall survival (OS).

RESULTS

Clinical features screening identified CRS, KRAS/NRAS/BRAF and liver lobe distribution as risk factors. Radiomics model, RQQM, combined model demonstrated higher AUC values compared to CRS and TBS model in training, internal and external validation cohorts (Delong-test P < 0.05). RQQM outperformed the radiomics model, but was slightly inferior to the combined model. Survival curves revealed statistically significant differences in 1-year DFS and 3-year OS for the RQQM (P < 0.001).

CONCLUSIONS

RQQM integrates both "quality" (radiomics) and "quantity" (TBS). The radiomics model is superior to the TBS model and has a greater impact on patient prognosis. In the absence of clinical data, RQQM, relying solely on imaging data, shows an advantage in predicting early recurrence after radical surgery for CRLM.

Deep learning-based multi-model prediction for disease-free survival status of patients with clear cell renal cell carcinoma after surgery: a multicenter cohort study

BACKGROUND

Although separate analysis of individual factor can somewhat improve the prognostic performance, integration of multimodal information into a single signature is necessary to stratify patients with clear cell renal cell carcinoma (ccRCC) for adjuvant therapy after surgery.

METHODS

A total of 414 patients with whole slide images, computed tomography images, and clinical data from three patient cohorts were retrospectively analyzed. The authors performed deep learning and machine learning algorithm to construct three single-modality prediction models for disease-free survival of ccRCC based on whole slide images, cell segmentation, and computed tomography images, respectively. A multimodel prediction signature (MMPS) for disease-free survival were further developed by combining three single-modality prediction models and tumor stage/grade system. Prognostic performance of the prognostic model was also verified in two independent validation cohorts.

RESULTS

Single-modality prediction models performed well in predicting the disease-free survival status of ccRCC. The MMPS achieved higher area under the curve value of 0.742, 0.917, and 0.900 in three independent patient cohorts, respectively. MMPS could distinguish patients with worse disease-free survival, with HR of 12.90 (95% CI: 2.443-68.120, P <0.0001), 11.10 (95% CI: 5.467-22.520, P <0.0001), and 8.27 (95% CI: 1.482-46.130, P <0.0001) in three different patient cohorts. In addition, MMPS outperformed single-modality prediction models and current clinical prognostic factors, which could also provide complements to current risk stratification for adjuvant therapy of ccRCC.

CONCLUSION

Our novel multimodel prediction analysis for disease-free survival exhibited significant improvements in prognostic prediction for patients with ccRCC. After further validation in multiple centers and regions, the multimodal system could be a potential practical tool for clinicians in the treatment for ccRCC patients.

Application research of radiomics in colorectal cancer: A bibliometric study

BACKGROUND

Radiomics has shown great potential in the clinical field of colorectal cancer (CRC). However, few bibliometric studies have systematically analyzed existing research in this field. The purpose of this study is to understand the current research status and future development directions of CRC.

METHODS

Search the English documents on the application of radiomics in the field of CRC research included in the Web of Science Core Collection from its establishment to October 2023. VOSviewer and CiteSpace software were used to conduct bibliometric and visual analysis of online publications related to countries/regions, authors, journals, references, and keywords in this field.

RESULTS

A total of 735 relevant documents published from Web of Science Core Collection to October 2023 were retrieved, and a total of 419 documents were obtained based on the screening criteria, including 376 articles and 43 reviews. The number of publications is increasing year by year. Among them, China publishes the most relevant documents (n = 238), which is much higher than Italy (n = 69) and the United States (n = 63). Tian Jie is the author with the most publications and citations (n = 17, citations = 2128), GE Healthcare is the most productive institution (n = 26), Frontiers in Oncology is the journal with the most publications (n = 60), and European Radiology is the most cited journal (n = 776). Hot spots for the application of radiomics in CRC include magnetic resonance, neoadjuvant chemoradiotherapy, survival, texture analysis, and machine learning. These directions are the current hot spots for the application of radiomics research in CRC and may be the direction of continued development in the future.

CONCLUSION

Through bibliometric analysis, the application of radiomics in CRC has been increasing year by year. The application of radiomics improves the accuracy of preoperative diagnosis, prediction, and prognosis of CRC. The results of bibliometrics analysis provide a valuable reference for the research direction of radiomics. However, radiomics still faces many challenges in the future, such as the single nature of the data source which may affect the comprehensiveness of the results. Future studies can further expand the data sources and build a multicenter public database to more comprehensively reflect the research status and development trend of CRC radiomics.

Deep learning or radiomics based on CT for predicting the response of gastric cancer to neoadjuvant chemotherapy: a meta-analysis and systematic review

Background

Artificial intelligence (AI) models, clinical models (CM), and the integrated model (IM) are utilized to evaluate the response to neoadjuvant chemotherapy (NACT) in patients diagnosed with gastric cancer.

Objective

The objective is to identify the diagnostic test of the AI model and to compare the accuracy of AI, CM, and IM through a comprehensive summary of head-to-head comparative studies.

Methods

PubMed, Web of Science, Cochrane Library, and Embase were systematically searched until September 5, 2023, to compile English language studies without regional restrictions. The quality of the included studies was evaluated using the Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2) criteria. Forest plots were utilized to illustrate the findings of diagnostic accuracy, while Hierarchical Summary Receiver Operating Characteristic curves were generated to estimate sensitivity (SEN) and specificity (SPE). Meta-regression was applied to analyze heterogeneity across the studies. To assess the presence of publication bias, Deeks' funnel plot and an asymmetry test were employed.

Results

A total of 9 studies, comprising 3313 patients, were included for the AI model, with 7 head-to-head comparative studies involving 2699 patients. Across the 9 studies, the pooled SEN for the AI model was 0.75 (95% confidence interval (CI): 0.66, 0.82), and SPE was 0.77 (95% CI: 0.69, 0.84). Meta-regression was conducted, revealing that the cut-off value, approach to predicting response, and gold standard might be sources of heterogeneity. In the head-to-head comparative studies, the pooled SEN for AI was 0.77 (95% CI: 0.69, 0.84) with SPE at 0.79 (95% CI: 0.70, 0.85). For CM, the pooled SEN was 0.67 (95% CI: 0.57, 0.77) with SPE at 0.59 (95% CI: 0.54, 0.64), while for IM, the pooled SEN was 0.83 (95% CI: 0.79, 0.86) with SPE at 0.69 (95% CI: 0.56, 0.79). Notably, there was no statistical difference, except that IM exhibited higher SEN than AI, while maintaining a similar level of SPE in pairwise comparisons. In the Receiver Operating Characteristic analysis subgroup, the CT-based Deep Learning (DL) subgroup, and the National Comprehensive Cancer Network (NCCN) guideline subgroup, the AI model exhibited higher SEN but lower SPE compared to the IM. Conversely, in the training cohort subgroup and the internal validation cohort subgroup, the AI model demonstrated lower SEN but higher SPE than the IM. The subgroup analysis underscored that factors such as the number of cohorts, cohort type, cut-off value, approach to predicting response, and choice of gold standard could impact the reliability and robustness of the results.

Conclusion

AI has demonstrated its viability as a tool for predicting the response of GC patients to NACT Furthermore, CT-based DL model in AI was sensitive to extract tumor features and predict the response. The results of subgroup analysis also supported the above conclusions. Large-scale rigorously designed diagnostic accuracy studies and head-to-head comparative studies are anticipated.

Systematic review registration

PROSPERO, CRD42022377030.

Machine Learning and Radiomics Analysis for Tumor Budding Prediction in Colorectal Liver Metastases Magnetic Resonance Imaging Assessment

PURPOSE

We aimed to assess the efficacy of machine learning and radiomics analysis using magnetic resonance imaging (MRI) with a hepatospecific contrast agent, in a pre-surgical setting, to predict tumor budding in liver metastases.

METHODS

Patients with MRI in a pre-surgical setting were retrospectively enrolled. Manual segmentation was made by means 3D Slicer image computing, and 851 radiomics features were extracted as median values using the PyRadiomics Python package. Balancing was performed and inter- and intraclass correlation coefficients were calculated to assess the between observer and within observer reproducibility of all radiomics extracted features. A Wilcoxon-Mann-Whitney nonparametric test and receiver operating characteristics (ROC) analysis were carried out. Balancing and feature selection procedures were performed. Linear and non-logistic regression models (LRM and NLRM) and different machine learning-based classifiers including decision tree (DT), k-nearest neighbor (KNN) and support vector machine (SVM) were considered.

RESULTS

The internal training set included 49 patients and 119 liver metastases. The validation cohort consisted of a total of 28 single lesion patients. The best single predictor to classify tumor budding was original_glcm_Idn obtained in the T1-W VIBE sequence arterial phase with an accuracy of 84%; wavelet_LLH_firstorder_10Percentile was obtained in the T1-W VIBE sequence portal phase with an accuracy of 92%; wavelet_HHL_glcm_MaximumProbability was obtained in the T1-W VIBE sequence hepatobiliary excretion phase with an accuracy of 88%; and wavelet_LLH_glcm_Imc1 was obtained in T2-W SPACE sequences with an accuracy of 88%. Considering the linear regression analysis, a statistically significant increase in accuracy to 96% was obtained using a linear weighted combination of 13 radiomic features extracted from the T1-W VIBE sequence arterial phase. Moreover, the best classifier was a KNN trained with the 13 radiomic features extracted from the arterial phase of the T1-W VIBE sequence, obtaining an accuracy of 95% and an AUC of 0.96. The validation set reached an accuracy of 94%, a sensitivity of 86% and a specificity of 95%.

CONCLUSIONS

Machine learning and radiomics analysis are promising tools in predicting tumor budding. Considering the linear regression analysis, there was a statistically significant increase in accuracy to 96% using a weighted linear combination of 13 radiomics features extracted from the arterial phase compared to a single radiomics feature.