Deep learning-enabled fluorescence imaging for oral cancer margin classification in preclinical models.

Area of Science:

• Biomedical Optics
• Medical Imaging
• Artificial Intelligence in Medicine

Background:

• Oral cancer surgery requires precise margin delineation to achieve complete tumor resection and preserve function.
• Current fluorescent optical imaging struggles with subsurface structures, leading to inadequate deep margins.
• Deep learning (DL) combined with structured light techniques offers potential for intraoperative margin assessment.

Purpose of the Study:

• To investigate a DL-enabled spatial frequency domain imaging (SFDI) system for subsurface depth quantification of fluorescent inclusions.
• To develop and validate DL models for predicting margin distance and fluorophore concentration in oral tumors.

Main Methods:

• A diffusion theory-based numerical simulation of SFDI was used for synthetic image generation for DL training.
• ResNet and U-Net convolutional neural networks were developed to predict margin distance and fluorophore concentration.
• Validation involved in silico, phantom, and ex vivo animal tissue datasets with fluorescent inclusions.

Main Results:

• The U-Net DL model predicted oral cancer depths with a mean error of 1.43 ± 1.84 mm and closest depths with 0.33 ± 0.31 mm error.
• In phantoms, subsurface depth was predicted with an error of 0.57 ± 0.38 mm.
• Ex vivo tissue analysis showed a closest distance prediction error of 0.59 ± 0.53 mm for inclusions up to 6 mm deep.

Conclusions:

• A DL-enabled SFDI system trained with in silico data shows significant promise for intraoperative margin assessment in oral cancer surgery.
• This approach can improve the accuracy of tumor resection by quantifying subsurface structures.
• Further development could enhance surgical precision and patient outcomes in oral cancer treatment.