Explainable liver tumor delineation in surgical specimens using hyperspectral imaging and deep learning

Area of Science:

• Medical imaging
• Computational pathology
• Surgical oncology

Background:

• Liver cancer recurrence after surgery is a major challenge due to residual malignant tissue.
• Visual inspection alone is insufficient for accurate tumor delineation during surgery.
• Hyperspectral imaging (HSI) offers a non-contact, label-free method to capture tissue spatial and spectral data.

Purpose of the Study:

• To investigate the feasibility of HSI for intra-operative liver tumor delineation on surgical specimens.
• To develop and evaluate a deep learning framework for HSI-based tumor identification.
• To identify key spectral channels for improved diagnostic accuracy and model interpretability.

Main Methods:

• A multi-task U-Net framework was developed for HSI data analysis.
• Measurements were conducted on 36 surgical specimens from 19 liver cancer patients.
• Pathological results served as the ground truth for model training and validation.
• A saliency-weighted channel selection method was employed to identify informative spectral channels.

Main Results:

• The U-Net framework achieved a sensitivity of 94.48% and a specificity of 87.22%.
• The developed method significantly outperformed the baseline Support Vector Machine (SVM) approach.
• A subset of 5 spectral channels provided comparable information to all 224 channels.
• Analysis of dominant channels suggested hemoglobin and bile content as potential markers.

Conclusions:

• HSI, coupled with a U-Net framework, is a feasible and effective tool for intra-operative liver tumor delineation.
• The proposed saliency-weighted channel selection method enhances efficiency and interpretability.
• Further investigation into hemoglobin and bile content differences could refine HSI-based diagnostics for liver cancer.