Dual-mode near-infrared multispectral imaging system equipped with deep learning models improves the identification of cancer foci in breast cancer specimens

For surgically resected breast cancer samples, it is challenging to perform specimen sampling by visual inspection, especially when the tumor bed shrinks after neoadjuvant therapy in breast cancer. In this study, we developed a dual-mode near-infrared multispectral imaging system (DNMIS) to overcome the human visual perceptual limitations and obtain richer sample tissue information by acquiring reflection and transmission images covering visible to NIR-II spectrum range (400–1700 nm). Additionally, we used artificial intelligence (AI) for segmentation of the rich multispectral data. We compared DNMIS with the conventional sampling methods, regular visual inspection and a cabinet X-ray imaging system, using data from 80 breast cancer specimens. DNMIS demonstrated better tissue contrast and eliminated the interference of surgical inks on the breast tissue surface, helping pathologists find the tumor area which is easy to be overlooked with visual inspection. Statistically, AI-powered DNMIS provided a higher tumor sensitivity (95.9% vs visual inspection 88.4% and X-rays 92.8%), especially for breast samples after neoadjuvant therapy (90.3% vs visual inspection 68.6% and X-rays 81.8%). We infer that DNMIS can improve the breast tumor specimen sampling work by helping pathologists avoid missing out tumor foci.