Supporting data for "CT Radiomics and Deep Learning Auto-segmentation in Epithelial Ovarian Carcinoma Treatment Response and Prognosis Evaluation"
Radiomics is a mathematical quantitative analysis that converts medical images into high-dimensional data by extracting large amounts of statistical features. Deep-learning(DL) based networks such as U-Net architecture have been increasingly explored in medical image segmentation to release radiologists and researchers from tedious annotation work, while no-new-Net (nnU-Net) resolves its weak generalization issues of semantic segmentation tasks in medical imaging. The series of original studies included in this thesis aimed to investigate and assess the clinical value of computed tomography (CT) radiomics and DL auto-segmentation in epithelial ovarian carcinoma(EOC) treatment response and prognosis evaluation.
Firstly, to predict platinum sensitivity which was determined by whether the patient relapsed within six months after platinum-based chemotherapy, the whole tumour volume was manually segmented on the baseline ceCT for feature extraction. Centres A-C were randomly divided into training and internal validation sets in 4:1 ratio. Centres D and E were assigned as independent external validation sets. Eleven radiomics features were selected and Extra Trees(ET) classifier was built across 10-fold stratified cross validation (SCV). The external validation set using centre D (n=44) had an area under curve(AUC) of 0.877; while centre E (n=51) had AUC of 0.845 in predicting platinum sensitivity.
Second study aimed to develop a DL algorithm in segmentation of omental metastases(OM) of EOC based on staging contrast-enhanced CT (ceCT) scans of EOC patients with OM from 6 institutions and to test its utility in recurrence detection. A cascade training configuration was used, which included 5-fold SCV with the image data in lower resolution followed by all images in full resolution. The nnU-Net segmentation model was built using OM on ceCT images of 627 patients (training n=478, internal validation n=74, external validation n=85). Our model achieved a Dice similarity coefficient (DSC) of 88.7±14.1% in an independent external validation (n=85) and a DSC performance of 61.7±22.7%in recurrence setting (n=80).
Third study evaluated the 3D volume change of OC and OM after neoadjuvant chemotherapy (NACT) using our pre-trained 3D nnU-Net auto-segmentation models, compared it with the Response Evaluation Criteria in Solid Tumors 1.1 (RECIST 1.1) and explored their prognostic value. Pre- and post-NACT ceCT scans of OC patients with OM were retrospectively collected and resampled (n=47). Their auto-segmentation was generated with their volume calculated automatically. NACT response according to RECIST 1.1 assessed by a radiologist was compared with the auto-generated volume. A macro AUC of 0.967 and micro AUC of 0.969 were achieved in predicting RECIST 1.1 with the total volume change. Pre-NACT volume of OM could predict platinum-sensitivity with an AUC of 0.800. Platinum-sensitivity, RECIST and total volume change were significant prognostic factors of overall survival . When they were enrolled into multi-variate Cox regression, only platinum-sensitivity was statistically significant. Total volume change was the only significant prognostic factor for progression free survival.
In summary, CT radiomics and deep learning auto-segmentation are valuable tools for predicting platinum sensitivity, detecting metastasis and recurrence, automating NACT response assessment, and evaluating prognosis in OC.
