Computer-Assisted Differentiation between Colon-Mesocolon and Retroperitoneum Using Hyperspectral Imaging (HSI) Technology.

Complete mesocolic excision (CME), which involves the adequate resection of the tumor-bearing colonic segment with "en bloc" removal of its mesocolon along embryological fascial planes is associated with superior oncological outcomes. However, CME presents a higher complication rate compared to non-CME resections due to a higher risk of vascular injury. Hyperspectral imaging (HSI) is a contrast-free optical imaging technology, which facilitates the quantitative imaging of physiological tissue parameters and the visualization of anatomical structures. This study evaluates the accuracy of HSI combined with deep learning (DL) to differentiate the colon and its mesenteric tissue from retroperitoneal tissue. In an animal study including 20 pig models, intraoperative hyperspectral images of the sigmoid colon, sigmoid mesentery, and retroperitoneum were recorded. A convolutional neural network (CNN) was trained to distinguish the two tissue classes using HSI data, validated with a leave-one-out cross-validation process. The overall recognition sensitivity of the tissues to be preserved (retroperitoneum) and the tissues to be resected (colon and mesentery) was 79.0 ± 21.0% and 86.0 ± 16.0%, respectively. Automatic classification based on HSI and CNNs is a promising tool to automatically, non-invasively, and objectively differentiate the colon and its mesentery from retroperitoneal tissue.

Automatic optical biopsy for colorectal cancer using hyperspectral imaging and artificial neural networks.

BACKGROUND: Intraoperative identification of cancerous tissue is fundamental during oncological surgical or endoscopic procedures. This relies on visual assessment supported by histopathological evaluation, implying a longer operative time. Hyperspectral imaging (HSI), a contrast-free and contactless imaging technology, provides spatially resolved spectroscopic analysis, with the potential to differentiate tissue at a cellular level. However, HSI produces "big data", which is impossible to directly interpret by clinicians. We hypothesize that advanced machine learning algorithms (convolutional neural networks-CNNs) can accurately detect colorectal cancer in HSI data. METHODS: In 34 patients undergoing colorectal resections for cancer, immediately after extraction, the specimen was opened, the tumor-bearing section was exposed and imaged using HSI. Cancer and normal mucosa were categorized from histopathology. A state-of-the-art CNN was developed to automatically detect regions of colorectal cancer in a hyperspectral image. Accuracy was validated with three levels of cross-validation (twofold, fivefold, and 15-fold). RESULTS: 32 patients had colorectal adenocarcinomas confirmed by histopathology (9 left, 11 right, 4 transverse colon, and 9 rectum). 6 patients had a local initial stage (T1-2) and 26 had a local advanced stage (T3-4). The cancer detection performance of the CNN using 15-fold cross-validation showed high sensitivity and specificity (87% and 90%, respectively) and a ROC-AUC score of 0.95 (considered outstanding). In the T1-2 group, the sensitivity and specificity were 89% and 90%, respectively, and in the T3-4 group, the sensitivity and specificity were 81% and 93%, respectively. CONCLUSIONS: Automatic colorectal cancer detection on fresh specimens using HSI, using a properly trained CNN is feasible and accurate, even with small datasets, regardless of the local tumor extension. In the near future, this approach may become a useful intraoperative tool during oncological endoscopic and surgical procedures, and may result in precise and non-destructive optical biopsies to support objective and consistent tumor-free resection margins.

Automatic Recognition of Colon and Esophagogastric Cancer with Machine Learning and Hyperspectral Imaging.

There are approximately 1.8 million diagnoses of colorectal cancer, 1 million diagnoses of stomach cancer, and 0.6 million diagnoses of esophageal cancer each year globally. An automatic computer-assisted diagnostic (CAD) tool to rapidly detect colorectal and esophagogastric cancer tissue in optical images would be hugely valuable to a surgeon during an intervention. Based on a colon dataset with 12 patients and an esophagogastric dataset of 10 patients, several state-of-the-art machine learning methods have been trained to detect cancer tissue using hyperspectral imaging (HSI), including Support Vector Machines (SVM) with radial basis function kernels, Multi-Layer Perceptrons (MLP) and 3D Convolutional Neural Networks (3DCNN). A leave-one-patient-out cross-validation (LOPOCV) with and without combining these sets was performed. The ROC-AUC score of the 3DCNN was slightly higher than the MLP and SVM with a difference of 0.04 AUC. The best performance was achieved with the 3DCNN for colon cancer and esophagogastric cancer detection with a high ROC-AUC of 0.93. The 3DCNN also achieved the best DICE scores of 0.49 and 0.41 on the colon and esophagogastric datasets, respectively. These scores were significantly improved using a patient-specific decision threshold to 0.58 and 0.51, respectively. This indicates that, in practical use, an HSI-based CAD system using an interactive decision threshold is likely to be valuable. Experiments were also performed to measure the benefits of combining the colorectal and esophagogastric datasets (22 patients), and this yielded significantly better results with the MLP and SVM models.

Deep Learning Analysis of In Vivo Hyperspectral Images for Automated Intraoperative Nerve Detection.

Nerves are critical structures that may be difficult to recognize during surgery. Inadvertent nerve injuries can have catastrophic consequences for the patient and lead to life-long pain and a reduced quality of life. Hyperspectral imaging (HSI) is a non-invasive technique combining photography with spectroscopy, allowing non-invasive intraoperative biological tissue property quantification. We show, for the first time, that HSI combined with deep learning allows nerves and other tissue types to be automatically recognized in in vivo hyperspectral images. An animal model was used, and eight anesthetized pigs underwent neck midline incisions, exposing several structures (nerve, artery, vein, muscle, fat, skin). State-of-the-art machine learning models were trained to recognize these tissue types in HSI data. The best model was a convolutional neural network (CNN), achieving an overall average sensitivity of 0.91 and a specificity of 1.0, validated with leave-one-patient-out cross-validation. For the nerve, the CNN achieved an average sensitivity of 0.76 and a specificity of 0.99. In conclusion, HSI combined with a CNN model is suitable for in vivo nerve recognition.

Automatic task recognition in a flexible endoscopy benchtop trainer with semi-supervised learning.

PURPOSE: Inexpensive benchtop training systems offer significant advantages to meet the increasing demand of training surgeons and gastroenterologists in flexible endoscopy. Established scoring systems exist, based on task duration and mistake evaluation. However, they require trained human raters, which limits broad and low-cost adoption. There is an unmet and important need to automate rating with machine learning. METHOD: We present a general and robust approach for recognizing training tasks from endoscopic training video, which consequently automates task duration computation. Our main technical novelty is to show the performance of state-of-the-art CNN-based approaches can be improved significantly with a novel semi-supervised learning approach, using both labelled and unlabelled videos. In the latter case, we assume only the task execution order is known a priori. RESULTS: Two video datasets are presented: the first has 19 videos recorded in examination conditions, where the participants complete their tasks in predetermined order. The second has 17 h of videos recorded in self-assessment conditions, where participants complete one or more tasks in any order. For the first dataset, we obtain a mean task duration estimation error of 3.65 s, with a mean task duration of 159 s ([Formula: see text] relative error). For the second dataset, we obtain a mean task duration estimation error of 3.67 s. We reduce an average of 5.63% in error to 3.67% thanks to our semi-supervised learning approach. CONCLUSION: This work is the first significant step forward to automate rating of flexible endoscopy students using a low-cost benchtop trainer. Thanks to our semi-supervised learning approach, we can scale easily to much larger unlabelled training datasets. The approach can also be used for other phase recognition tasks.