New Technique Combines ML with SWIR Fluorescence Imaging for Precise Surgical Tumor Removal

Recent advances in imaging and biochemical technologies have improved a surgeon's ability to distinguish between tumors and healthy tissue. One such technique that enables this distinction is "fluorescence-guided surgery" (FGS). A new study proposes a method for classifying healthy and tumor cells using an intensity-independent approach. This method combines machine learning with short-wave infrared (SWIR) fluorescence imaging to precisely detect the boundaries of tumors.

FGS involves staining the patient's tissue with a dye that emits infrared light when irradiated with a special light source. The dye selectively binds to the surface of tumor cells, enabling the detection of the location and extent of the tumor based on the emitted lightwaves. However, most FGS-based methods rely on the absolute intensity of the infrared emissions to differentiate pixels corresponding to tumors. This approach is problematic since intensity is influenced by lighting conditions, camera setup, dye quantity, and staining duration. Therefore, intensity-based classification can lead to inaccurate interpretation.

The new technique developed by researchers at the University College London (London, UK) involves capturing multispectral SWIR images of the dyed tissue, rather than relying solely on measuring the total intensity over one wavelength. To achieve this, the team sequentially placed six different wavelength frequency (color) filters in front of their SWIR optical system and registered six measurements for each pixel. By doing this, the researchers were able to create spectral profiles for each type of pixel, including background, healthy tissue, and tumor. Subsequently, they trained seven machine learning models to accurately identify these spectral profiles in multispectral SWIR images.

The research team conducted in vivo training and validation of the models using SWIR images of an aggressive type of neuroblastoma in a lab model. They also evaluated various normalization techniques to make pixel classification independent of absolute intensity and dependent only on the pixel's spectral profile. The study involved testing seven machine learning models, with the top-performing model achieving a remarkable per-pixel classification accuracy of 97.5%. Specifically, the accuracies for tumor, healthy, and background pixels were 97.1%, 93.5%, and 99.2%, respectively.

In addition, the model's results were found to be highly robust against variations in imaging conditions due to the normalization of the spectral profiles. This is desirable for clinical applications because testing of new imaging technologies is typically done in ideal conditions that are not reflective of the real-world clinical setting. Based on their findings, the research team is optimistic about the potential of this methodology. They believe that conducting a pilot study to implement it in human patients could lead to significant advancements in the field of FGS.

Multispectral FGS has the potential to go beyond the current study's scope. It can be used to remove unwanted reflections and surgical or background lights from images, as well as offer noninvasive ways of measuring lipid content and oxygen saturation. Multispectral systems also allow for the simultaneous use of multiple fluorescent dyes with different emission characteristics since the signals from each dye can be untangled from the total measurements based on their spectral profiles. This multiple dye approach can target multiple aspects of a disease, providing surgeons with even more information. Future studies will undoubtedly explore the full potential of multispectral FGS, unlocking doors to more effective surgical procedures for treating cancer and other illnesses.

Related Links:
University College London