1431P - Assessing pathological complete response to neoadjuvant chemotherapy combined with immunotherapy in esophageal squamous cell carcinoma: A deep learning approach with voxel-level radiomics

Background

Neoadjuvant chemotherapy combined with immunotherapy (nCIT) is emerging as a promising strategy for treating patients with esophageal cancer. We aimed to develop a novel, voxel-level radiomics deep learning model to predict the pathological complete response (pCR) to nCIT in esophageal squamous cell carcinoma (ESCC).

Methods

We included 745 ESCC patients from three institutions, comprising a training set (N=500, institution 1), test-set-1 (N=86, institution 1), test-set-2 (N=34, institution 2), and test-set-3 (N=125, institution 3). All patients underwent nCIT and received a pathological evaluation post-surgery. A total of 104 distinct radiomic features were extracted from the tumor areas in CT images prior to nCIT. Features demonstrating predictive power with an area under the receiver operating characteristic curve (AUROC) exceeding 0.6 were retained. Each selected radiomics feature map, along with the original image, was input into a vision transformer model as a separate channel. The Shapley Additive exPlanations method was employed to visualize feature importance and identify predictive regions within the tumor, enhancing the model’s interpretability. For baseline comparison, we also constructed a clinical model using only clinical factors, and compared it against traditional radiomics and ResNet models.

Results

The voxel-level radiomics deep learning model achieved an AUC of 0.825 with an accuracy of 76.2% in test-set-1, 0.751 and 68.5% in test-set-2, and 0.819 with 78.1% in test-set-3, respectively. It demonstrated superior predictive performance compared to the clinical model, traditional radiomics model, and ResNet model.

Conclusions

This innovative and noninvasive voxel-level radiomics deep learning approach offers an efficient and accurate approach of predicting treatment response to nCIT in ESCC, potentially benefiting clinical decision-making.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

Z. Zhang.

Funding

Has not received any funding.

Disclosure

All authors have declared no conflicts of interest.