Combining radiomics and deep features in positron emission tomography improves accuracy in tumour response to neo-adjuvant chemoradiotherapy in oesophageal squamous cell carcinoma

Abstract

Purpose: This study examined whether radiomics and deep learning are predictive of tumour response in patients with oesophageal squamous cell carcinoma (OSCC) treated by neoadjuvant chemoradiotherapy (nCRT) and surgery. Material and methods: Ninety-five patients (65 training and 30 validation cohorts) who had undergone pre-treatment fluorodeoxyglucose positron emission tomography (18F-FDG-PET) studies were included and classified as those who achieved complete pathological response (pCR) and those who did not. The primary tumours were segmented using a fixed threshold approach and radiomics features extracted to construct a hand-crafted radiomics (HCR) signature. The contoured volume of interest (VOI) and the PET images were input into a fully convolutional neural network (CNN) for volumetric medical imaging (VNet) for the prediction of pCR. A third model was constructed combining both the radiomics and deep features. Results: pCR was achieved in 36 (37.9%) patients. The HCR signature achieved an area under the receiver operating characteristic curve (AUC) of 0.787. We observed substantial overfitting in our deep learning model and instead, the CNN generated VOI of the primary tumour and 8 stable deep features. A single deep feature was shown to be predictive of pCR with an AUC of 0.741. The model was improved when both the HCR and deep features were combined (AUC 0.839). Conclusion: HCR features provide incremental value to deep learning for predicting pCR in OSCC patients undergoing nCRT. Our approach not only has the potential to provide a personalised management plan for these patients but also demonstrates a novel approach in tackling small medical imaging datasets using deep learning.