CT-based deep learning may help predict chemoradiotherapy response in oesophageal squamous cell carcinoma

Researchers from the Sun Yat-Sen University Cancer Centre (SYSUCC) and the University of Hong Kong (HKU) have developed a novel deep learning model based on patients’ CT scans – the ResNet50-support vector machine (RN-SVM) model – with improved efficiency and accuracy in predicting response to neoadjuvant chemoradiotherapy (nCRT) in oesophageal squamous cell carcinoma (ESCC) compared with a handcrafted radiomics model and a clinical model.

“Our study’s findings suggested that transfer learning would be a suitable alternative for medical imaging in situations where data might be insufficient,” said the researchers. “Further radiogenomic analysis in our study provided important insights into the biological mechanisms of treatment resistance to nCRT in ESCC.” [Radiother Oncol 2021;154:6-13]

The retrospective study included data from 231 patients (mean age, 60 years; male, 83.1 percent) with ESCC treated with nCRT and surgery between April 2007 and December 2018. Records of 161 patients (cohort 1, model training; mean age, 58 years) and 70 patients (cohort 2, model testing; mean age, 64 years) were extracted from SYSUCC and HKU, respectively.

Baseline characteristics, in terms of pathological complete response (pCR) rate (46.0 percent vs 44.3 percent; p=0.93), cN stage (p=0.48), clinical stage (p=0.22) and histological grading (p=0.73), were comparable between cohort 1 and cohort 2.

Six pretrained convolutional neural networks (CNNs) were used to extract deep learning features from baseline enhanced CT scans of cohort 1. The last fully connected layer at the top of each network was first removed. The maximum value of each layer of feature map on the networks was then extracted with global max pooling and transforming into raw values.

Handcrafted radiomic features, developed from CT scans of patients in cohort 1, were computed automatically from regions of interest (ROIs) manually segmented from radiologists. Defined features with or without wavelet filtration were extracted.

A clinical model based on clinical factors were also built for baseline comparison, with classification probability regarded as the radiological score.

Of the six radiological models developed, the ResNet50-SVM model achieved the best classification performance, and was superior to the radiomic model and clinical model (area under curve [AUC], 0.805 vs 0.725 vs 0.508; accuracy, 77.1 percent vs 67.1 percent vs 47.1 percent). The positive predictive value and negative predictive value of the ResNet50-SVM model was 70.3 percent and 84.8 percent, respectively.

The tumour area and peritumoural region on CT images were found to be valuable in feature pattern extraction.

A combination of deep learning and radiomics features, however, failed to show an improvement in prediction performance, with AUC of 0.799. Likewise, incorporation of clinical factors into radiological models did not improve the performance.

Results of radiogenomic analysis of 385 enriched gene sets suggested a potential link between radiological scores and biological processes involving the WNT signalling pathway, as well as tumour microenvironmental components such as proteoglycan, extracellular matrix (ECM), immune cells and hypoxia.

“We developed and validated a model using transfer learning technique to perform pretherapeutic assessment of nCRT response in ESCC. The RN-SVM model offers the advantages of better performance and a more efficient procedure without manual tumour contouring compared with the handcrafted radiomics method,” the investigators concluded.