Modern visceral surgery is often performed through small incisions. Compared to open surgery, these minimally invasive interventions result in smaller scars, fewer complications and a quicker recovery. While to the patients benefit, it has the drawback of limiting the physician’s perception largely to that of visual feedback through a camera mounted on a rod lens: the laparoscope. Conventional laparoscopes are limited by “imitating” the human eye. Multispectral cameras remove this arbitrary restriction of recording only red, green and blue colors. Instead, they capture many specific bands of light. Although these could help characterize important indications such as ischemia and early stage adenoma, the lack of powerful digital image processing prevents realizing the technique’s full potential.
The primary objective of this thesis was to pioneer fluent functional multispectral imaging (MSI) in laparoscopy. The main technical obstacles were: (1) The lack of image analysis concepts that provide both high accuracy and speed. (2) Multispectral image recording is slow, typically ranging from seconds to minutes. (3) Obtaining a quantitative ground truth for the measurements is hard or even impossible.
To overcome these hurdles and enable functional laparoscopy, for the first time in this field physical models are combined with powerful machine learning techniques. The physical model is employed to create highly accurate simulations, which in turn teach the algorithm to rapidly relate multispectral pixels to underlying functional changes. To reduce the domain shift introduced by learning from simulations, a novel transfer learning approach automatically adapts generic simulations to match almost arbitrary
recordings of visceral tissue. In combination with the only available video-rate capable multispectral sensor, the method pioneers fluent perfusion monitoring with MSI. This system was carefully tested in a multistage process, involving in silico quantitative evaluations, tissue phantoms and a porcine study. Clinical applicability was ensured through in-patient recordings in the context of partial nephrectomy; in these, the novel system characterized ischemia live during the intervention. Verified against a fluorescence reference, the results indicate that fluent, non-invasive ischemia detection and monitoring is now possible.
In conclusion, this thesis presents the first multispectral laparoscope capable of videorate functional analysis. The system was successfully evaluated in in-patient trials, and future work should be directed towards evaluation of the system in a larger study. Due to the broad applicability and the large potential clinical benefit of the presented functional estimation approach, I am confident the descendants of this system are an integral part
of the next generation OR.