This multi-centre investigation explored the effectiveness of a novel nomogram integrating CECT radiomics and DL characteristics with independent imaging features to distinguish retroperitoneal lipomas from WDLPS by predicting the MDM2 amplification status. The performance of the DLRN surpassed that of the RS, DLRS, and clinical models, highlighting the added utility of integrating radiomics and DL features with established clinical variables for differentiating retroperitoneal lipomas from WDLPS. The DLRN also exhibited a high level of calibration and yielded substantial net benefits, indicating its potential as a dependable and efficient tool to help clinicians design individualised treatment plans.
Accurately distinguishing retroperitoneal lipomas from WDLPS is crucial for effective treatment planning. Several imaging features, including presence of thick septa, increased non-fat tissue content, and tumour size, have been reported as differentiation criteria [18–20, 27]. However, distinction solely on the basis of visual inspection can be challenging due to the variations in appearance caused by necrotic foci, infarctions, inflammation, or non-adipose tissue [31]. Previous studies on systematic radiological interpretations showed a relatively limited level of reproducibility among observers, with kappa agreement scores ranging from 0.17 to 0.42 [22, 23, 32]. In our study, the clinico-radiological model did not distinguish between WDLPS and retroperitoneal lipomas with high accuracy (AUC = 0.700, accuracy = 0.683 in the external validation set). Since radiologist experience plays a crucial role in interpretations of imaging characteristics, clinical experience could have influenced the trustworthiness and objectivity of the judgements, indicating why the diagnostic efficiency of the clinical model was lower than that of models based on radiomics and DL signatures.
MDM2 amplification is regarded as one of the initial events in the pathogenesis of WDLPS [12, 13], and is not typically observed in benign lipomas. Radiomics is an effective and non-invasive tool to predict gene expression, as shown in studies related to non-small cell lung cancer [33], glioblastoma [34], breast cancer [35], colorectal cancer [36], and renal clear cell carcinomas [37], and can offer diagnostic and treatment benefits in cancer interventions. In our study, most of the 10 radiomics features chosen to construct an RS to identify WDLPS and retroperitoneal lipomas by predicting MDM2 amplification were texture features, which are not readily interpretable but included more intricate information about intra-tumoural heterogeneity and histologic status and exhibited greater sensitivity for gene expression prediction.
DL algorithms, which can capture in-depth information within images through specific tasks in the hidden layers of neural networks, have shown potential in the detection and management of soft tissue sarcomas (STS). Navarro et al. [38] investigated the correlation between a multimodal DL signature and the grading of pathological tumours in soft tissue sarcomas patients, and found that the model showed promise in prognosticating the outcomes of STS patients, with an AUC value of 0.760 in the test cohort. In our study, we used the transfer learning technique to extract representative DL features from pre-trained CNNs, and the DLRS with 13 DL features and 3 HCR features showed encouraging performance in distinguishing between WDLPS and lipomas (AUC = 0.861 in the test cohort). DL and radiomics employ distinct techniques for picture evaluation, yielding complementary, rather than complimentary, features. The DLRS demonstrated better performance compared than the RS in identifying WDLPS and retroperitoneal lipomas and achieved higher values for both the AUC and accuracy, further demonstrating the additional value of DL features.
Several radiomics and DL studies have distinguished liposarcoma from lipoma with relatively high accuracy, showing the potential of radiomics models for improving clinical decision-making. Thornhill et al. [39] incorporated a comprehensive range of liposarcoma subtypes into their model, encompassing examples like myxoid liposarcoma and dedifferentiated liposarcoma, which exhibit distinct radiological characteristics and can be easily discriminated from lipoma by experienced radiologists. However, our study focused exclusively on the two most challenging tumour types to distinguish (WDLPS and lipoma), making our dataset both clinically relevant and highly demanding. In addition, prior studies did not address WDLPS/atypical lipomatous tumours occurring in the retroperitoneal and extremities independently. In contrast, this study solely focused on distinguishing retroperitoneal WDLPS from lipomas, making it more targeted and specific. We chose to focus on the retroperitoneal regions instead of the extremities due to the following reasons. First, radiomics and DL research specifically targeting retroperitoneal liposarcomas is limited. Thus, our study would prove beneficial to both clinicians and scientists involved in addressing this specific disease within the realm of personalised medicine. Second, retroperitoneal liposarcoma is associated with increased surgical difficulty and a more unfavourable prognosis than extremity liposarcoma, possesses distinct natural characteristics, and is managed differently. Lastly, retroperitoneal WDLPS are more likely to undergo dedifferentiation than those that occur in the extremities. Early and accurate diagnosis, as well as complete resection, can improve prognosis.
Several limitations of this study require consideration. First, even with the strict exclusion criteria, selection bias could not be ruled out because of the retrospective nature of this investigation. Therefore, well-planned prospective radiomics and DL trials are required to validate the proposed imaging biomarkers. Second, the ROIs were manually delineated, resulting in variations across different observers. Therefore, semi-automated methods for ROI delineation could be beneficial. Third, the integration of emerging imaging technologies with radiomics and DL (such as positron-emission tomography-CT and dynamic contrast-enhanced MRI) represents future directions of exploration. Fourth, this investigation enrolled a limited study population because of the infrequent occurrence of WDLPS and retroperitoneal lipoma. Finally, CT devices varied across different participating centres in this study, but the ComBat harmonisation method was employed to mitigate batch effects and enhance model accuracy.