The incidence of PCNSL has been steadily rising over the last 30 years.[17] Due to the lack of specific clinical manifestations, there is a certain misdiagnosis rate in imaging diagnosis.[18, 19] The treatment method of PCNSL is different from that of conventional tumors. The treatment of PCNSL is mainly combined with highdose methotrexate (HD-MTX) chemotherapy and whole brain radiotherapy (WBRT) (45 Gy).[9] Hence, early diagnosis is crucial to avoid unnecessary surgical resection.[3, 20] Furthermore, early diagnosis and follow-up examination after therapy play a crucial role in the prognosis of PCNSL.[21] Previous studies suggested that among PCNSL patients, low-risk of early progression patients may benefit from MTX alone, while those at high risk of early progression may require additional chemotherapy and/or WBRT.[22, 23] The radiomics model we constructed was able to grade tumors by analyzing MRI images (ADC, DWI and T1-CE).
Previous studies have shown that Ki-67 could indicate the tumor aggressiveness and it is an important predictor for disease prognosis.[24–26] Ki-67 is an independent prognostic factor in determining post-metastatic overall survival (PMOS) in certain malignancies.[27] It is closely related to the increase of cell number and volume, and this will restrict the diffusion of water molecules which increase the signal of DWI. At the same time, high levels of Ki-67 also represent a strong invasive ability of the tumor, that is, the ability to infiltrate and destroy blood vessels. PCNSL lesions usually have cuffing infiltration around blood vessels which destroys vascular endothelial cells. In this case, the contrast agent seeps from the damaged blood-brain barrier around the tumor. As a result, hypo-vascular PCNSL is often markedly homogeneously enhanced on T1-CE images. Moreover, the higher the level of Ki-67, the more active the proliferation of tumor cells. The greater the damage to the blood-brain barrier, the more obvious the tumor enhanced. Therefore, to find whether ADC, DWI and T1-CE parameters are related to Ki-67 levels in primary CNS lymphoma is important for interpreting the biological significance of medical images. In this study, we found that rDWImean, rADCmean and rT1-CEmean were significantly associated with Ki-67.
ADC can detect the diffusion of water molecules in lesions to reflect the microstructure of lesions. A significant negative correlation between Ki-67 and ADC parameters was observed in breast cancer[28] and liver cancer.[29] In the study of Schob et al.,[30] Ki-67 also showed a significant negative correlation with ADC fractions in PCNSL. This is consistent with our results. The results of this study showed that rT1-CEmean was significantly correlated with Ki-67, and there was a significant difference between the high proliferation group and the low proliferation group. No relevant studies have reported the relationship between rT1-CEmean and Ki-67 previously. However, it has been demonstrated in nasopharyngeal carcinoma that parameters of T1-CE MRI may be associated with disease prognosis.[31]
More and more studies have focused on radiomics for the prediction of disease treatment response,[32] prognosis prediction,[33] exploring the relationship between radiomics features and biological features.[34] Based on the above research results, we constructed a radiomics model that can predict the proliferation level of PCNSL. In order to ensure the reliability of the prediction model, this study excluded variable factors by three steps. Firstly, in order to reduce the adverse effects caused by the singular sample data, we standardized and normalized the data. Then radiomics features were extracted from each patient's ADC, DWI, and T1-CE images, respectively. The second step is to exclude the differences between radiologists in describing ROI, and select omics features with high stability (ICC ≥ 0.8) for the next feature screening. The last step is to adjust the threshold of feature correlation analysis and use F test for feature screening.
ROC curve showed that the model constructed by all sequences combined, sequence combination (DWI + T1-CE) and radiomics features in T1-CE images had good diagnostic performance. The AUC of all sequence combination validation sets was 0.869 (95% CI 0.772–0.9657), which had the best diagnostic performance. The reason is that radiomics features based on multiple sequences can independently and perfectly describe tumor information and are significantly correlated with Ki-67.
Radiomics can extract highthroughput quantitative image features to characterize lesion features.[35] These features include geometry, signal intensity and histogram, and image texture features to reflect the level of tumor proliferation and may explain the difference between the high and low proliferation groups of PCNSL.[36] Most of the features cannot be visually obtained by the naked eye, so the number is far more than the imaging features observed by naked eye of radiologists. Therefore, the model's description of the tumor may be more perfect, which has been confirmed in the study of Xia et al..[37] Previous studies have constructed radiomics models with good diagnostic performance for the diagnosis of typical or atypical PCNSL and the differential diagnosis of PCNSL and glioma, helping clinicians to diagnose in patients who cannot undergo stereotactic biopsy.[38, 39] The radiomics model constructed in this study may help to predict the tumor proliferation level of patients with stereotactic biopsy contraindications and provide more reliable decision support for tumor treatment and reexamine after treatment. This model could be benefit to avoid the optimal treatment time window missing for patients with biopsy contraindications, and to enable personalized medicine by adjusting the dose of chemotherapy and HBRT.
The decision analysis curve has unique advantages in analyzing the clinical value of the model. We found that in the test set, the model achieved a net clinical benefit after grouping patients for Ki-67 compared to no grouping patients. Therefore, radiomic models can predict the disease prognosis and provide clinical treatment decision support.
Our study has some limitations. First of all, due to the low incidence, although we collected patients from many hospitals, this study still included relatively less cases. More patients with PCNSL needed to be included and perform independent external validation. In addition, the ability of the scientific research platform used in this study to extract features is limited, and feature extraction needs to be improved in the future. Finally, due to the lack of specific clinical manifestations of PCNSL, this study did not construct a nomogram in combination with clinical factors, and the past studies have shown that nomograms perform significantly better than routinely used clinical tumor staging, tumor size and clinical models.[40] Therefore, in future work, clinical factors of PCNSL need to be extracted and combined with radiomics features to show nomograms.