Machine Learning-Based Pathomics Signature for Perineural Invasion in Colorectal Cancer

BACKGROUND Perineural invasion (PNI) is strongly associated with poor clinical outcomes in colorectal cancer (CRC). However, no machine learning diagnostic model based on pathomics has been established for PNI detection in CRC. To address this issue, we sought to construct a predictive model for PNI grounded in pathological features to enhance diagnostic efficiency. MATERIAL AND METHODS We analyzed hematoxylin and eosin-stained histopathological slides from the CRC tissues retrospectively. Segmentation of the acquired images was conducted via CellProfiler, an automated pipeline supporting the extraction of morphological features. To optimize feature selection, we applied the LASSO algorithm, followed by multiple machine learning models to develop diagnostic classifiers for PNI. Furthermore, we investigated the clinicopathological significance of PNI, including its association with T stage, lymph node metastasis, lymphovascular invasion, and molecular biomarkers. RESULTS We used 430 CRC surgical resection slides for training, testing, and external validation. A total of 615 histopathological features were extracted, and 10 of them were screened by LASSO to construct diagnostic models for PNI. The models demonstrated robust predictive performance across all cohorts. LightGBM achieved the highest diagnostic accuracy, yielding AUCs of 0.996 (95% CI: 0.991-1.000, training), 0.935 (95% CI: 0.888-0.978, testing), and 0.918 (95% CI: 0.861-0.967, external validation). Patients with CRC with PNI exhibited higher T stage, increased lymph node metastasis, and more frequent lymphovascular invasion. CONCLUSIONS The LightGBM model, based on histopathological features, can improve the diagnostic efficiency of PNI. CRC with PNI is associated with poor prognosis.