Previous studies have shown that 18F-FDG PET/CT radiomics performed well in predicting the prognosis of various malignancies. The newly developed PET/CT radiomic signature was a powerful predictor of gastric cancer survival . Radiomics features of baseline PET/CT images provide complementary prognostic information for nasopharyngeal carcinoma compared with the use of clinical parameters alone . This method was also advantageous to predict the prognosis of lung cancer, breast cancer and other tumors [35–37]. As for colorectal cancer, a few studies demonstrated that FDG PET radiomic held potential towards the improved prediction of clinical outcome in stage IV patients of colorectal cancer and locally-advanced rectal cancer [38, 39]. The explosive researches on the prognostic value of PET/CT-based radiomics methods for the total colorectal cancer were rare, especially for stage III. A study on the National Cancer Data Base (NCDB) showed that CRC patients with stage III accounted for approximately one-third of all stages . Moreover, the 5-year survival rate of this largest proportion of patients was subjected to a large difference in the survival outcomes [41, 42]. Therefore, evaluating the prognosis of stage III colorectal cancer separately and intervening as early as possible according to different individual patients to reduce the risk of recurrence and metastasis is necessary. In this study, we developed an original model to predict the prognosis of CRC patients and further experimented on stage III patients by 18F-FDG PET/CT radiomics.
We originally investigated prognosis models for patients ranging from stage I to IV in the primary experiment. The model M1 trained by a combination of features of all three modalities outperformed other models with a C-index of 0.780 [95% CI 0.634-0.877]. The K-M curves indicated that high-risk and low-risk patient groups could be separated by our model effectively (P < 0.0001). For model M1, CA199 was the most important feature. It means that this cancer antigen marker CA199 contributed most to the outcome of the prognostic prediction in model M1. The result is consistent with previous studies  that CA199 is a key prognostic biomarker. It should also be noted that the contribution of imaging features was irreplaceable, although they only accounted for 13.3% of the contribution. Both PET and CT features were important and irreplicable in radiomics analysis because they both had positive importance scores, which suggests these features positively contributed to the model accuracy. Experimental results in Manuscript Fig. 3A verified that the model constructed with multimodalities (C-index 0.780) outperformed the models built with PET (C-index 0.592) or CT (C-index 0.755) alone on D-1~4. Similar trend can be identified on D-3 with Fig. 3B.
We also focused on analyzing models for patients with stage III, because the 5-year survival rate is unsatisfactory, though radical surgery and adjuvant chemotherapy were routinely performed. The prediction of prognosis is valuable for supporting individualized treatment. The C-index of M2 was 0.820 [95% CI 0.676-0.900], which means it holds a great potential value of prognostic prediction in colorectal cancer. Its performance was also superior to single modality or double modality models. K-M curves of M2 illustrate the model could significantly separate high-risk and low-risk patient group. For model M2, PET-Wavelet-LLH-glszm-ZV was the most important feature in the predictive model, which means the texture information quantified by this PET feature successfully captured the heterogeneity of colorectal tumour towards prognostic prediction. It was because PET images could provide information not only about the metabolism of the tumor, but also about the total load of the tumor. For further interpretation of this PET feature, we conducted correlation analysis, and found that this PET features positively correlated with 40% MTV and TLG (p<0.05). CA199 which contributes most in M1 only made up 14% of all feature contributions.
Moreover, CA199, lymph nodes and CT-Log-sigma-5.0-3D-glszm-SAE were three features identified both in M1 and M2. The feature importance analysis showed that clinical features played the most vital roles in the prognosis of CRC patients of all stages while radiomics features made more contribution when predicting the prognosis of CRC patients with stage III. The case study also demonstrated that features with greater contribution could help the model to overcome the negative impact caused by single features then rectify the prediction. Thereby, it is reasonable to believe that the combination of clinical characteristics and imaging characteristics of 18F-FDG metabolism is more convincing than any single modality model.
We reduced the risk of overfitting through reducing the number of features and employed cross-validation in feature selection. Firstly, we reduced the risk of overfitting by strictly controlling the number of features, as the reduced number of features leaded to the decrease of the number of required parameters inside machine learning models, thus minimizing the risk of overfitting . According to the guideline for radiomics studies , we reduced the number of features to less than 1/10 of sample sizes. Secondly, 50 times five-fold cross-validation was deployed during the feature selection on the training dataset to reduce the risk of overfitting . By selecting features on the rotating training instead of a fixed training set, we effectively minimize the risk of overfitting on a fixed proportion of data. Thirdly, we evaluated the risk of overfitting by comparing the performance of the model on training and testing datasets in independent validation. Supplementary Table S1 shows the difference between training and testing C-index was less than 0.03 in both experiments, which suggests the risk of overfitting was properly handled.
This study was partly limited by its retrospective design and relatively modest sample sizes. We will continue to collect more patients who meet the criteria and attempt to conduct prospective studies to further validate our models. We look forward to further randomized controlled trials in the future on the significance and importance of 18F-FDG PET/CT imaging omics in the diagnosis and treatment of colorectal cancer. We will investigate the effect of spatial resolution of PET/CT images on the parameters of radiomic feature extraction.