Optimisation and evaluation of the random forest model in the efficacy prediction of chemoradiotherapy for advanced cervical cancer based on radiomics signature from high-resolution T2 weighted images

Abstract

Purpose: Our objective was to establish a random forest model and to evaluate its predictive capability of the treatment effect of neoadjuvant chemotherapy-radiation therapy.

Methods: This retrospective study included 82 patients with locally advanced cervical cancer who underwent scanning from March 2013 to May 2018. The random forest model was established and optimised based on the open source toolkit scikit-learn. Byoptimising of the number of decision trees in the random forest, the criteria for selecting the final partition index and the minimum number of samples partitioned by each node, the performance of random forest in the prediction of the treatment effect of neoadjuvant chemotherapy-radiation therapy on advanced cervical cancer (> IIb) was evaluated.

Results: The number of decision trees in the random forests influenced the model performance. When the number of decision trees was set to 10, 25, 40, 55, 70, 85 and 100, the performance of random forest model exhibited an increasing trend first and then a decreasing one. The criteria for the selection of final partition index showed significant effects on the generation of decision trees. The Gini index demonstrated a better effect compared with information gain index. The area under the receiver operating curve for Gini index attained a value of 0.917.

Conclusion: The random forest model showed potential in predicting the treatment effect of neoadjuvant chemotherapy-radiation therapy based on high-resolution T2WIs for advanced cervical cancer (> IIb).

Keywords: Cervical cancer; Chemoradiotherapy; Radiomics; Random forest; T2-weighted image.

Fig. 1

Recruitment pathway for patients in this study. Ninety-eight patients received pathological diagnosis for cervical cancer were included initially. Sixteen patients were excluded according to the exclusion criteria, and a total of 82 patients were eventually included in the study. CR complete response, PR partial response

Fig. 2

Radiomics workflow of model construction. a MR images segmentation. First, the tumour was segmented manually on the sagittal image, and then ITK-SNAP was used for 3D volume reconstruction. b Radiomic feature extraction. According to the segmentation image, a total of 106 radiomics parameters of 3 types were extracted from each set of images. c Radiomic Feature selection. After the preprocessing of Wavelet Filtering and Laplacian of Gaussian, the characteristic parameters were selected and classified by decision-making tree. d Model establishment. The diagnostic efficacy of the radiomics model was evaluated by ROC analysis

Fig. 3

The development process diagram of random forest model. The training set is divided into N training subsets, and each subset generates a decision tree. A total of N decision trees are generated, and the n decision trees are assembled together. This process is the construction process of the random forest model

Fig. 4

a The relationship between the number of decision trees in random forest and the performance of the model. When the number of decision trees was set to 10, 25, 40, 55, 70, 85 and 100, respectively, the performance of random forest model shows a trend of rising first and then declining, and its inflection point appeared at 70. b Area under curve (AUC) at different final partition index. The Gini index demonstrated a better effect compared with Information gain index. For Gini index model, the AUC is 0.864. c AUC at different number of sub-samples (Nsub) randomly sampled for each decision tree in training. The best performance was acquired when Nsub was equal to the log₂N which AUC is 0.871. d Area under curve (AUC) at different minimum number of samples (Nmin) partitioned by each node. The best performance was acquired when Nmin was equal to 6 which AUC is 0.914