2087 - Integration of MRI Radiomics and Host Inflammatory and Nutritional Factors for Predicting Pathological Complete Response after Chemoradiotherapy in Esophageal Squamous Cell Carcinoma

03:00pm - 04:00pm ET

Poster Hall - Exhibit Hall A

Screen: 22

POSTER

Presenter(s)

Koichi Hayano, MD, PhD - Chiba University Hospital, Chiba, Chiba

K. Hayano, A. Hirata, T. Tochigi, Y. Kurata, A. Kakimoto, and G. Ohira; Department of Frontier Surgery, Chiba University Graduate School of Medicine, Chiba, Japan

Purpose/Objective(s):

This study aimed to develop a model integrating MRI-based radiomics and host inflammatory and nutritional factors to predict pathological complete response (pCR) and stratify recurrence risk in patients with advanced esophageal squamous cell carcinoma (ESCC).

Materials/Methods:

Seventy patients with ESCC who underwent CRT followed by curative surgery were retrospectively investigated. Seventy-six radiomics features were extracted from pre-treatment diffusion-weighted MRI (b=0 and 1000 s/mm²) using a radiomics software (PixSpace Inc., Fukuoka, Japan). Gray-level co-occurrence matrix (GLCM) features were calculated using two intensity-rescaling approaches: global min–max scaling and localized ±3SD normalization. Host-related variables included white blood cell count with differential, platelet count, C-reactive protein (CRP), prognostic nutritional index (PNI), liver-to-spleen ratio (LSR), and body mass index (BMI). Feature screening was performed by statistical software (SAS Institute Inc., Cary, NC, USA), and the top four MRI radiomics features and top four host factors were selected separately. All possible combinations within each group were evaluated using five-fold cross-validation. Model performance was assessed by cross-validated area under the receiver operating characteristic curve (AUC) using automated machine learning software (Sony Network Communications Inc., Tokyo, Japan).

Results:

pCR was achieved in 15 patients (21.4%). In the host-domain analysis, the combination of monocyte count, CRP, and LSR achieved the highest predictive performance with an AUC of 0.765, while the optimal radiomics-only model, consisting of three features (b1000 GLCM-Correlation, b0 GLCM-Autocorrelation [global], and b0 GLCM-Autocorrelation [±3SD]), demonstrated a superior AUC of 0.846. The hybrid model integrating these top radiomics features with the host monocyte count yielded the highest overall performance, reaching an AUC of 0.853. Furthermore, patients were stratified into two groups based on the median predicted probability of pCR (0.3727). Kaplan–Meier analysis demonstrated that the group with a high-probability of pCR had significantly better recurrence-free survival compared to the low-probability group (p=0.03, log-rank).

Conclusion:

An integrated model combining MRI radiomics and the host monocyte count demonstrated robust performance for predicting pCR and stratifying postoperative recurrence risk in ESCC. Our results might provide an important insight into selecting the optimal therapeutic strategy for the treatment of advanced ESCC.