A non-invasive multi-phase CT classifier for predicting pre-treatment enlarged lymph node types in colorectal cancer

Abstract

Background

Colorectal cancer (CRC) with benign lymph node enlargement (BLNE) (>1 cm) is often associated with better long-term prognosis and favorable outcomes in immunotherapy. However, lymph node enlargement (LNE) can mislead clinicians into considering metastatic lymph node enlargement (MLNE), potentially resulting in misguided therapeutic decisions in unnecessary neoadjuvant therapy and extended lymphadenectomy. This, ultimately, can lead to overtreatment, increasing the risk of postoperative complications and tumor recurrence. Thus, developing a pre-treatment multimodal CT radiomics-based model to assess LNE status is essential.

Methods

A total of 319 pre-treatment multimodal CT images of CRC patients with LNE were retrospectively collected from 2015 to 2020 as a development cohort. Additionally, 111 multimodal CT images from 2020 to 2022 were prospectively collected as a validation cohort. Tumor and LNE regions of interest were manually segmented, and 40 patients were randomly re-outlined by another radiologist to extract radiomics features. The intragroup correlation coefficient was calculated to assess the reproducibility of the radiomics features. Following feature screening, multiple predictive models were constructed, including tumor and lymph node models for individual modalities (Tumor_N, _A, _V; Ln_{N, A, V}; Ln, lymph node; N, non-contrast phase; A, arterial phase; V, venous phase), along with 15 models combining multiple modalities. The predictive performance of these models was assessed using area under the receiver operating characteristic curve (AUROC) and precision-recall curve (AUPRC), along with sensitivity, specificity, and accuracy.

Findings

After validation with the prospective cohort, Tumor_N and Ln_A demonstrated the best predictive performance among single modalities, with AUROC values of 0.626 and 0.781, respectively. Among all models, Ln_NAV exhibited the highest predictive performance, achieving AUROC and AUPRC values of 0.820 and 0.883, respectively, with a sensitivity of 0.708, specificity of 0.848, and overall accuracy of 0.766.

Interpretation

Radiomics, as a non-invasive and quantitative approach, can reflect underlying physiopathological information. The incorporation of a multimodal radiomics model yielded excellent performance in predicting pre-treatment LNE status, particularly for BLNE, with a specificity of 0.848. This approach can provide valuable guidance for clinical treatment strategies.