¹⁸F‑FDG PET/CT based radiomics features improve prediction of prognosis: multiple machine learning algorithms and multimodality applications for multiple myeloma

The second most prevalent hematologic malignancy is MM. It is a plasma cell disorder characterized by aberrant monoclonal plasma cell (PC) proliferation in bone marrow (BM) [1]. It develops from monoclonal gammopathy of undetermined significance (MGUS), and undergoes an intermediate phase known as smoldering MM (SMM), before advancing to active MM [2].

At present, MM is diagnosed using laboratory- and image-based assessments. The most widely utilized imaging methods for MM diagnosis are CT and PET/CT. The bone and extraosseous symptoms of MM can be evaluated using PET/CT in terms of their presence, size, and metabolic activity [3]. 20% of newly diagnosed MM patients still have a dismal prognosis, despite recent improvements in new therapies for the survival of MM patients [4, 5]. Therefore, predicting MM patient prognosis and treating them with precision can potentially enhance patients' survival in the future [6]. In the clinics, cytogenetic examinations using bone marrow biopsy and aspiration are necessary for the prognostic stratification and precise treatment of MM patients [7]. However, this invasive procedure is often painful for the patients, and samples are difficult to obtain. There are times when an intrusive biopsy is ineffective on the first try and requires subsequent biopsies. Additionally, due to the heterogeneity of the obtained biopsy material, it may not be typical of the whole malignancy parts and only represents a tiny fraction of it.

Due to the limitations of conventional approaches outlined above, finding noninvasive ways of prognostic prediction among MM patients is a topic of significant research interest. There is a recent rise in using machine learning (ML) algorithms to predict patient prognosis based on radiomics features. This is a noninvasive and accurate mean of prognostic stratification, and it was previously reported in MM patients [8]. Schenone et al. employed ML algorithms to accurately stratify the outcome of autologous transplanted MM patients based on CT radiomics features. This, in turn, assisted in designing proper personalized treatments for individual patients [9]. Although the Li et al. study analyzed an ML model constructed from MRI-based radiomics features, in combination with a clinical model for MM, MRI is typically not the examination of choice for MM patients, and certain shortcomings still exist in this regard [3, 10]. Given these evidences, it is imperative to identify novel biomarkers for the accurate prediction of MM patient prognosis. Based on prior research, radiomics-based prognostic signatures have great potential in stratifying MM patients as either low- or high-risk. This information is crucial, particularly, for high-risk patients, who, with advanced therapy, may experience enhanced outcomes.

Furthermore, a multimodal application of ML has emerged in recent years that integrates radiomics features from PET, CT, and so on to construct ML models. This approach facilitates a closer examination of the entire intra-heterogeneous tumor, rather than the limited information gathered from unimodal medical images [11]. Haider et al. used 5 machine learning algorithms and multimodal model integrating both PET and CT features to analyze the relationship between oropharyngeal squamous cell carcinoma and human papillomavirus, with an area under the AUC of 0.78 [12]. Although there are related studies that predicted MM patient prognosis based on PET/CT imaging radiomics features. However, to date, there are no reports on employing the multimodal radiomics features of PET and CT and multiple ML algorithms to predict MM patient prognosis. Hence, herein, we utilized the aforementioned method to study prognosis and to enhance personalized patient care in order to better MM patient outcome [13, 14].

At present, the role of ¹⁸F-FDG PET/CT in diagnostics and response evaluation criteria among MM patients has reached an extremely significant level of evidence for clinical decision making, prognosis determination, and treatment response evaluation [32‐34]. However, only a few prior studies have examined the prognostic values of radiomics features in MM. Using the Cox regression model, Yang et al. confirmed that the radiomics profiles of bone marrow MRI exhibit obvious correlation with MM patient OS. Moreover, the predictive performance of radiomics-based signature is far superior to the traditional clinical model [10]. Ludivine Morvan et al. provided a novel radiomics feature selection protocol for ¹⁸F-FDG PET-based radiomics in MM patients, and they emphasized the advantages of employing image-based characteristics (including, textural profiles) for disease progression estimation [35]. In addition to the prognostic investigations, ¹⁸F-FDG PET/CT-based radiomics have been used in other different application of areas in MM. In comparison to human specialists, the radiomics model showed a significant improvement in diagnostic capacity. For example, the PET radiomics measure is quite effective in differentiating between bone metastases and vertebral MM [36]. However, to our knowledge, this report is the first to utilize CT and PET-based radiomics features, as well as combined clinical parameters implemented with multiple ML methods, to predict prognosis of MM patients. We employed a total of 10 image types and 8 features types in this study, and fully explored the radiomics features of PET/CT in MM patients in detail.

The bootstrap method was used in this study. Bootstrap is a common statistical method that uses repeated resampling of data samples to generate larger sample sets, thereby avoiding the limitation of small sample sizes. Similar bootstrap procedures were implemented in studies of Yilong Huang et al. [37], Wen L et al. [38], and Mostafa Nazari et al. [39]The significant differences between the clinical models and the radiomics models were compared using two statistical methods [40]. Using multiple feature screening methods, we retained the features with clinical interpretability while eliminating the multicollinearity among features.

Based on our analysis, the late RISS, ISS staging, high-risk genetic status, anemia, high serum globulin level, and low albumin levels were strongly associated with shorter PFS, which corroborates with published clinical trials and findings within clinical practice [34, 41]. PET-based features original_shape_MinorAxisLength, original_shape_LeastAxisLength_PET, MTV, sMTV demonstrated relevancy of PFS in the step1 for MM progress in this study. However, they were excluded from analysis during the feature selection step 2 due to the multicollinearity because the coefficients became 0 at the optimal penalty parameter. Our rationale was that these features simultaneously reflected high tumor burden and late tumor stage, and thus, depicted the magnitudes of the tumors.

The exclusion of CT-based features such as wavelet.HHH_gldm_SmallDependenceEmphasis and wavelet.HLH_gldm_SmallDependenceEmphasisCT was due to their extraction process from the same feature type of radiomics feature and the same type of filter of image, with only different combinations generated by using high-pass and low-pass filters in each of the three dimensions. Thus, these features also exhibit high collinearity and were therefore excluded from the analysis.

Other factors that are known to impact MM prognosis are high level LDH, hypercalcemia, renal insufficiency, and baseline treatment arms, TLG and sTLG. However, in this study, these factors did not reach significance likely due to the limited data and unintentional selection bias. It is also possible that these factors have less prognostic significance, compared to the other factors identified in this report, which is similar to the findings of the Yang Li et al. and Bastien Jamet et al. studies [10‐13].

Using RSF and GBM algorithms, we further enhanced PFS accuracy in various modalities, and improved the overall clinical decision-making process. However, the overall GB-COX and CoxBoost algorithms accuracies did not improve significantly, compared to the multivariate COX regression models, yet they significantly improved the clinical decision-making ability in the combinations of PET + CLI and CT + CLI. Unlike the boosting-based algorithm,the decision tree-based algorithm RSF/GBM was more prone to overfitting producing higher differences between the bootstrap training sets and the validation sets. However, the SVRC algorithm achieved the lowest predictive performance in our systematic evaluation likely due to our feature selection steps, as such, this requires further exploration in future investigations [35, 42, 43]. Furthermore, in this investigation, we observed that the PET-based radiomics features were better in predicting MM patient prognosis, compared to CT-based radiomics features. We also demonstrated that integrating the radiomics signature with patient clinical profile greatly enhanced the prognostic predictive ability of our model. Based on this evidence, the radiomics signature is critical for estimating patient prognosis. It not only provides prognostic efficiency of bone marrow PET/CT radiomics in MM patients, but also has potential for clinical risk classification. This report offers an essential supplementary reference for radiomics-based prognostic models as we compared numerous ML methods for PFS estimation of MM patients. This large-scale comparison is beneficial for the accurate selection of ML methods for radiomics-based PFS estimation.

There were several restrictions on this work. First of all, because it was a single-center retrospective study, there could have been accidental bias in the patient selection process. As a result, our conclusions are not representative of the general population. Secondly, owing to a relatively small patient population and absence of an external groups for validation, we employed the bootstrap technique for model assessment to circumvent the issue of limited sample population. We recommend that future investigations assess the proposed prognostic model among a large multicenter-recruited patient population, and validate the results in an external independent validation cohort.

¹⁸F-FDG PET/CT based radiomics models implemented with machine learning algorithms can significantly improve the clinical prediction of progress and increased clinical benefits providing prospects for clinical prognostic stratification for precision treatment as well as new research areas.