We proposed a volume-related feature selection approach to reduce dimensionality and identify those parameters that are relevant for HL characterization. It is known that a multitude of radiomics parameters are strongly correlated one to another, implying high redundancy that affects radiomics models’ performance. Dimensionality reduction and feature selection are crucial mandatory steps before any modeling (Sollini et al., 2020; Park et al., 2019). In fact, as recommended, an adequate ratio between the number of features and the number of patients should be preserved. Additionally, lesion morphology and size in lymphoma patients are similar within and among patients. Therefore, radiomics features unbiased from the volume and shape descriptors need to be identified in radiomics applications in lymphoma. Our approach allowed us to select a set of features (ranging from 2 to 15) to be used for model building. The main advantage of the proposed method for features reduction and selection relies on the concept that a fingerprint, comprising volume-related and non-volume related features, is able to represent all the lesions of a patient.
In view of the multisite disease, identification of the lymphoma lesion to be processed for radiomics analyses is crucial. We proposed an innovative approach for radiomic-wise lesion similarity assessment to provide the evidence for target selection. Conventionally, the largest lesion or the one with the highest FDG uptake is used but the rationale for this method has never been supported by any evidence. We demonstrated that the lesions within a patient may show different grades of similarity. Intra-patient lesion similarity within R/R patients was higher compared to non-R/R. Interestingly, intra-patient lesion similarity in the R/R dataset was confirmed also when extra-nodal lesions were included in the analysis. In the non-R/R group, the addition of extra-nodal lesions to nodal ones had a minor effect on similarity.
The non-relapsing/refractory (non-R/R) group is a homogeneous subset of patients; it included all cases before treatment initiation, and they were included in one single institution. On the other hand, the relapsing/refractory (R/R) group included patients treated with several lines of treatment coming from different centers. These two scenarios allowed us to explore lesion similarity in two opposite situations. Furthermore, we aimed at identifying a radiomic fingerprint that could be representative of HL lesions irrespective of all the variables, with the long-term goal of wide application of the fingerprint among different centers. We, indeed, found that the R/R, even if it could be expected to be more heterogeneous, resulted to have higher intra-patient similarity as compared to non-R/R.
It should be acknowledged that the number of observations (i.e., lesions) may have partially affected these findings. However, we did not expect to provide definitive results but to propose a methodological framework for future investigations. Indeed, our “proof-of-concept” approach resulted encouraging for further development for response prediction. We foresee the necessity of research in this direction since among the available studies, the bias related to a significant disproportion between the patient groups (responders vs non-responders being the latter less than 10% of the whole cohort (Milgrom et al., 2019)) may have significantly influenced the results. Additionally, the intra-patient lesion similarity in non-R/R patients was scarce even when a higher number of lesions were analyzed (Fig. 3a), suggesting that this group of patients was intrinsically more heterogenous. This finding was expected since non-R/R HL, naïve from any treatment, included patients who later on experienced long-term response, relapse, and refractory disease; therefore, it was the most heterogenous group. Conversely, R/R patients may be biologically more homogeneous, since treatments might result in resistant clones’ selection. Moreover, the non-neoplastic cells of tumor microenvironment have been claimed as one of the main determinants responsible for pathogenesis and progression of HL (Mottok & Steidl, 2018; Calabretta et al., 2019). Infiltration of the tumor microenvironment by CD68+ and CD163+ macrophages, Treg and CD4+ T cells (especially with Th2 phenotype), and high CD4/CD8 ratio is associated to the emergence of resistance to conventional therapy, and a worse prognosis. Additional factors that dysregulate tumor microenvironment promote a vicious loop between malignant cells and the components of the tumormicroenvironment stimulating resistance to treatment and disease progression. These factors include the recruitment of tumor-associated macrophages, the secretion of cytokines with macrophage chemotactic activity reinforced by the reactive cells, the activation of fibroblasts promoted by molecules secreted by malignant cells, the expression of surface antigens (e.g., CD30L, CD40L) by inflammatory cells that act as survival signals for the neoplastic cells, and the aberrant activation of signaling pathways (e.g., NF-κB, PI3K) (Karantanos et al., 2017) promote a vicious loop between malignant cells and the components of the tumor microenvironment stimulating resistance to treatment and disease progression (Karantanos et al., 2017). Of note, evidence suggests that [18F]FDG uptake is more likely related to elements of microenvironment rather than malignant HL cells (Gillessen et al., 2020; Barrington & Mikhaeel, 2014; Shim et al., 2009). Accordingly, our findings are in line with the fact that heterogeneity of the tumor microenvironment in naïve patients is more pronounced than that of R/R patients. Therefore, our results support the need for development of a radiomics fingerprint in a large cohort of naïve patients. Essentially, in this analysis, we explored and developed a framework for radiomics analysis in lymphoma. Simultaneous presence of many lesions is a typical finding in lymphoma, and recent data on molecular profiles suggest lesions’ heterogeneity (Spina et al., 2018; Banerjee, 2011).
The question, related to the choice of which and/or how many lesions, which guide the disease, and need to be processed, is unresolved. In image mining studies, one possible approach to address this issue is the choice of the largest and/or the most metabolically active lesion, as for conventional image analysis and adopted by previous studies (Ben Bouallègue et al., 2017; Tatsumi et al., 2019). However, large heterogeneous lesions (often necrotic or with multiple uptake peaks) may underestimate the volume (El-Galaly et al., 2018; Carles et al., 2017) and influence texture measurements. On the other hand, all the lesions could be considered for radiomics analysis. As demonstrated in the present study, enriching the analysis through the use of the information derived from all lesions improved the classification performance. Results of the classifier using the largest lesion were not satisfactory (accuracy = 60%), but the small sample size prevents any speculation about their reliability. Conversely, the RUBTE provided promising results when all lesions were used for the analysis, similarly to the previous investigations (Lue et al., 2019; Ganeshan et al., 2017; Parvez et al., 2018; Mayerhoefer et al., 2019). Unlike in the study by Milgrom et al., the authors found the mediastinal lesion-derived features could predict patient outcome, while features extracted from all lymphoma sites did not predict refractory disease (Milgrom et al., 2019). Overall, segmentation or annotation of all lesions is time-consuming and could hardly be implemented into the clinical routine practice. Therefore, suitable trade-off considering the number of cases at hand and the needed predictive power is necessary.
HL typically involves more than one site, and lesions different in size may co-exist. We found that the PCA-derived information mapping volume data outnumbered non-volume ones in almost all cases, with the exception of fingerprint_One—the one built on the largest lesion. Therefore, the huge variability in lesions’ size within patients required more covariates (i.e., features) to characterize the lesions and to be inclusive for all lesions. Our results are encouraging for exploring the proposed framework in larger multicenter trials. We foresee a replication study to confirm our data. Secondly, we propose that future radiomics investigations on lymphoma have to rely on the radiomics features derived from all the lesions of a patient. The approach we developed may be applied also for solid tumor studies if multiple lesions are present, in order to understand from which lesion (primary, secondary or all) to extract the features for modeling and predictions.
Some limitations should be acknowledged including the retrospective design and sample size, even if the involvement of more centers conferred strength to results. We pooled features extracted from images acquired using different scanners (Orlhac et al., 2018). On the other hand, we did not search for feature cutoff in the analysis. Moreover, we had previously demonstrated that scanners and image postprocessing did not affect final results (Kirienko et al., 2018). Additionally, to test our research hypothesis, we evaluated the lesions within the same patient; therefore, the scanning protocol and postprocessing were consistent among lesions. We developed a fingerprint for each group of patients. Obviously, the development of one fingerprint representative for all lesions regardless the site (nodal or extra-nodal) and the dataset (non-R/R or R/R) would be the ultimate goal. However, the primary aim of this preliminary analysis was to test if really radiomics differed in non-R/R and R/R (i.e., define a methodological framework to demonstrate the potential predictive value of radiomics in HL). Background activity may affect segmentation and, consequently, feature calculation. Nonetheless, the introduction of extra-nodal lesions improved the silhouette index in non-R/R (Fig. 4c) and R/R datasets (Fig. 5c), respectively. We could speculate that, irrespective of the possible issues in extra-nodal lesions segmentation, lesion texture did not result in higher inhomogeneity. However, these results should be confirmed in larger datasets, since in our cohort only 27 patients had extra-nodal lesions. When lesions were adjacent to areas of high physiological uptake, we avoided to include those lesions for radiomics analysis in order not to introduce a bias in lesion segmentation. We operated that choice since we expected it to be more robust and generalizable for future studies. Additionally, we decided to avoid considering diffuse uptake disease in bone, spleen, and liver in the present analysis in order not to introduce a potential bias in image interpretation since diffuse uptake may have been related to both disease infiltration and functional activation. Lastly, within the inter-patient analysis, we compared patient populations in two different settings—naïve patients at staging (non-R/R) and patients candidate to immunotherapy who failed several lines of treatment (R/R). This choice was based on the expectation that the class of non-R/R HL accounting for patients that did not recurred after at least 4 years from first-line treatment completion (i.e., cured HL) would have differed the most from the class of R/R.