Up to 22% of polyps are missed during colonoscopy and current solutions relay on improving the endoscopic image or using AI. 

Microwave-based colonoscopy (MWC) is a disruptive new solution that has proved to be safe and feasible in a human pilot study. MWC can differentiate polyps from normal tissue based on their different dielectric properties and consists on three elements: the acquisitor (an add-on ring-shaped device provided with microwave antennas), the analyzer (external unit) and the software (embedded in the analyzer). The software analyses the signals received and provides with an output (sound). Unfortunately, there are no available  microwave signal databases for colorectal polyps that can be used for training the algorithm and we need more representative examples to improve the performance. For this reason, we designed a dedicated phantom using an ex-vivo porcine colon that included six synthetic polyps selected from a set of models with different morphologies, including flat polyps 2 mm in height and 25, 10 or 7 mm in diameter, as well as polyps 10 mm in diameter and 5 or 10 mm in height. Performance metrics were evaluated after the explorations. We used the optical colonoscopy video as ground truth. The optical video from colonoscopy was examined and temporally segmented into sets of consecutive and homogeneous frames. 

We performed a total of 24 colonoscopies with a total of 144 polyps and 32,441 frames. A convolutional neural network (CNN) was trained on these data, and each detection was labelled as  TP, TN, FP and FN. The CNN was then validated using six new trajectories that included 36 flat polyps. The sensitivity and specificity for polyp detection were globally 80.5% and 75.0%. respectively. Sensitivity increased up to 93.3% for flat polyps.

MWC has the potential for detecting colon flat polyps and can be trained in dedicated synthetic phantoms.