Intraoperative margin assessment in colorectal cancer via AI-enhanced multiphoton microscopy.

Abstract

Significance: Identifying residual tumors at the resection margins is crucial for individuals undergoing colorectal cancer (CRC) surgery. Accurate margin assessment can help minimize residual tumor while avoiding unnecessary removal of healthy tissue.

Aim: This study uses label-free multiphoton microscopy to visualize collagen architecture and develops a feature-enhanced deep-learning model to distinguish CRC from margin-adjacent normal tissue.

Approach: Margin-adjacent normal and carcinoma tissues were imaged using multiphoton microscopy, and 142 quantitative collagen features were extracted from SHG images. We propose PVT-Enhanced, a feature-fusion model that integrates handcrafted collagen features with deep image features derived from PVTv2-b5 through an attention-based fusion strategy. Its performance was compared with LASSO-LR, three convolutional neural networks, and image-only PVTv2-b5.

Results: SHG imaging showed relatively coherent fiber orientation, longer and more continuous collagen fibers, and higher fiber density in margin-adjacent normal tissue, whereas carcinoma tissue showed more disorganized orientation, shorter and more fragmented fibers, and lower fiber density. The PVT-Enhanced model achieved an accuracy of 86.9%, an AUC of 0.923, and a specificity of 96.4%. Only 5 of 137 margin-adjacent normal images were misclassified as CRC.

Conclusions: Our findings suggest that the PVT-Enhanced framework combined with SHG imaging allows accurate, rapid, and label-free tissue characterization that may support intraoperative margin assessment in CRC surgery.