The future of capsule endoscopy in clinical practice: from diagnostic to therapeutic experimental prototype capsules

SonoCAIT

The group of Steward et al. [34] has developed a proof-of-concept capsule for an ultrasound (US)-mediated targeted drug delivery (UmTDD), named SonoCAIT. The prototype capsule is suitable for drug delivery in a specific location, utilising US to release drugs and/or to increase drug uptake through sonoporation. SonoCAIT has a pill-like shape with the dimensions of 10 mm diameter and 30 mm length. An US transducer, a drug delivery channel, a vision module, and a multi-channel external tether are integrated into the capsule. The vision module consists of a CMOS camera and circuit board with four white LEDs. The camera is cylindrical with resolution of 220 × 224 pixels. The aim of SonoCAIT is to deliver drugs to the wall of the GI tract. One example of a therapeutic preparation is drug-filled microbubbles (MBs). When these reach the target zone, they must be released in close proximity to the wall, where the drugs can then be released by US. That means the US focus and MBs have to be directed towards the same target. In vitro experiments showed that SonoCAIT achieved enhancement of drug uptake.

A capsule for haemostasis utilising an inflated balloon

An inflatable prototype capsule for haemostasis in the GI tract based on balloon tamponade effect has been proposed by the group of Leung et al. [35]. The capsule consists of three segments linked with flexible joints. These segments are a gas generation chamber with a length of 13 mm, an acid injector with a length of 35 mm, and a circuit box with a length of 12 mm. The capsule has a diameter of 14 mm, and it is enclosed in a silicone balloon. The balloon inflates at a bleeding lesion and achieves haemostasis by the tamponade effect. The inflation of the balloon is achieved by an acid injection into a gas generation chamber filled with base powder. The amount of infused acid controls the pressure and the volume of the silicone balloon in order to suit the variation of the diameter and texture of the intestine. The inflation of the balloon is capable of producing appropriate pressure to the bowel wall to anchor the capsule steadily in the position of the bleeding. Ex vivo experiments for the evaluation of the appropriate pressure and in vivo experiments of bleeding in the small intestine of a pig were conducted, showing that the proposed capsule is able to achieve haemorrhage control in the lower GI.

A wireless spiral capsule robot with modular structure

Detection and treatment of GI pathologies are two functionalities, which need different hardware integration, and the co-existence of both in the same capsule is limited by its size. Thus, the group of Guo et al. [36] has proposed a wireless spiral capsule robot with modular structure driven by the external magnetic field. The capsule robot consists of two robots with modular structure with a mechanism to combine and separate them. It consists of a guide robot and an auxiliary robot with helical diversion grooves. The guide robot has a length of 39.4 mm and a diameter of 12.5 mm. The auxiliary robot has a length of 33.8 mm and a diameter of 12.5 mm. In the two robots, permanent magnets have been placed in the centre and magnetized in the radial direction. Under the same external magnetic field, generated by the three-axis Helmholtz coils, the motion of each capsule is relative to the other. After the inspection of the guide module robot, the treatment module robot can be swallowed and then dock with the guide module robot, thus saving the long time normally required to reposition the target location.

Blue and red light capsule for photodynamic therapy of Helicobacter pylori

The group of Tortora et al. [37] investigate the efficiency of LEDs in specific wavelengths needed for the therapy of infection from Helicobacter pylori. Based on their measurements, two easy-to-swallow capsule devices have been developed. The first is a capsule for research purposes, which has a more performant battery that is not permitted in clinical examination. The second is a preindustrial capsule that integrates certified modules for clinical examination. The capsules integrate eight LEDs placed in an electronic board along with a magnetic switch and a battery. Two versions of the capsule have been constructed based on the emitting wavelength. One emits red light only at 625 nm and the other blue light at 405 nm. The capsules have a diameter of 14 mm and length of 27 mm.

Soft-magnet-based drug-delivery module for the ALICE System

A module for drug delivery that is combined with the ALICE [25] system has been presented by the group of Le et al. [38]. The drug delivery module consists of two ring-type soft magnets and a simple plastic hinge. The ring-type magnets are axially magnetised, attracting each other and keeping the drug enclosed inside the module. The ALICE system provides controlled navigation of the integrated ALICE and drug delivery module to investigate and accurately infuse the drug to the lesion area. The drug-delivery module is opened by the repulsive force between the two radially magnetised soft-magnetic rings, when the axial magnetisation of the rings stops. The rings are demagnetised and a strong pulsating magnetic field in a radial direction is applied, and the drug enclosed in the module is released. Then, the two rings are axially magnetised again, attracting each other, and thanks to the plastic hinge the drug-delivery module is returned to its initial shape. The integrated drug delivery module with ALICE system has a diameter of 12 mm and a length of 33 mm. The proposed drug delivery module with ALICE system has been tested in vitro.

Wireless capsule endoscope with holding mechanism for medication release

The group of Woods et al. [39] have presented a concept of a microrobot with a medication release and a drug infusion mechanism for targeted drug delivery. The microrobot has been designed with resistance to natural peristalsis, deploying a holding mechanism enabling the microrobot to localise a pathological area of interest in the GI tract. Then, a needle is positioned in this area and delivers a 1-ml dose of medication. The needle has the ability to be placed in a 360° scale, while simultaneously maintaining a diametrically opposite relationship with the holding mechanism. This feature guarantees the penetration of the GI tract wall by the needle. The holding mechanism utilises a single micromotor to open and close two legs. The legs stretch and hold the microrobot in the region of interest inside the GI tract.

Blue light therapy capsule for Helicobacter pylori

A blue light emission capsule for the therapy of Helicobacter pylori has been proposed by the group of Li et al. [40]. A module for pH sensing and measuring is used to differentiate locations and evaluate the digestive function by monitoring the pH values of the GI tract. The optical source of the capsule consists of eight blue LEDs which emit blue light for treatment according to a pre-set range of pH values. Also, the capsule consists of a low power-consumption microcontroller for the processing of the pH signal and a wireless communication module for the transmission of the measured pH values to an external receiver. The proposed capsule has a diameter of 11.5 mm and a length of 22 mm.

Capsule with a coil actuation mechanism for drug release

The group of Beccani et al. [41] has designed a prototype Magnetic Drug Delivery Capsule (MDDC). A coil actuation mechanism is used by the capsule for the release of the drug. The capsule consists of a coil, a magnet, and a drug chamber in which matching magnets are placed. The drug chamber remains attached to the body of the capsule by the generated attraction between the magnets. The coil is able to produce the appropriate force for the magnets to be repulsed. Then, the chamber opens, and the drug is released in the specific area of the GI tract. The diameter of the capsule is 13 mm and the length is 30 mm. The capsule has a weight of 12 g, and the drug chamber can store to 2.4 ml of drug.

Drug delivery capsule in a specified location using magnets

A smart capsule for location-specific drug release in the GI tract has been presented by the group of Yu et al. [42]. The proposed capsule has length of 26 mm and diameter of 9 mm. The specific location for the release of the drug is determined by an implanted or externally worn permanent magnet. The capsule is activated when it is close to the magnet. Then, a reed switch closes and the capacitor is discharged by a nichrome wire. A nylon fuse melts while the cap is opening and the drug is released. The drug delivery procedure was tested in vitro showing the locomotion of the capsule next to the permanent magnet.

Blood detection methods

Computer-based abnormality detection systems have proved especially useful in detecting and locating the origin of obscure gastrointestinal bleeding (OGIB), which is defined as chronic bleeding from a source not found after traditional (wired) endoscopy. Most of the proposed systems concentrate on features associated with the colour of the image, as blood has a distinct red hue. Colour is the most important aspect that differentiates bleeding and non-bleeding regions. However, in some cases edge pixels, such as those found in intestinal folds, and bleeding pixels share similar dark hues, which means that traditional algorithms often mistake edge pixels for bleeding pixels. Fu et al. [44] and Usman et al. [45] used a technique to remove these edge pixels aiming to enhance bleeding detection. Usman et al. used [45] the transformation of an image to Hue saturation value (HSV) colour space to extract image features, while Fu et al. [44] grouped pixels adaptively into uniformly-coloured segments based on colour and location, with a procedure called superpixel segmentation. The feature used from each superpixel for bleeding detection is the ratio of red to green, blue, or the sum of red, green, and blue intensities in RGB (Red, Green, Blue) colour space.

Yuan et al. [46] considered quantised colour histograms as features for blood detection, using a technique named Bag of Words (BoW). This technique represents each image with a set of “visual” words, corresponding to image segments. In the second stage, the bleeding regions are localised using a saliency map that indicates regions of importance within the image, estimated based on colour information from various colour spaces. A classifier fusion algorithm to detect the bleeding frames and localise the bleeding area was proposed by Deeba et al. [47]. It combines the results of two classifiers trained using first-order statistical features extracted from RGB and HSV colour spaces, which include mean, standard deviation, entropy, skew, and energy. Furthermore, Ghosh et al. [48] noted that the blue (B) component of RGB does not carry any valuable information for the discrimination between bleeding and non-bleeding zones. Instead, a composite colour component obtained by dividing the green (G) with the red (R) components, i.e. G/R, was found to be more informative. First-order statistical features were extracted from G/R and were used as the input to an SVM classifier for the detection of bleeding frames.

As well as colour features, texture and shape features have been also considered for the discrimination of bleeding frames. In the study conducted by Hu et al. [49], a geometric image feature, called local-contrast-enhanced higher-order local auto-correlation (LCEHLAC), was utilised along with an image pre-processing method for a non-linear conversion model of the HSV colour space. A methodology combining colour histogram features and automatically extracted features was proposed by Jia et al. [50]. The former are extracted from CIE-Lab colour space, and the latter are extracted using a deep (multi-layer) neural network architecture, called the Convolutional Neural Network (CNN).

A summary of the aforementioned blood detection methods along with their results in various datasets is provided in Figure 1. It can be noticed that in all cases the performance metrics are high, exceeding 90%. The different methods cannot be directly compared to each other because they have been tested on different datasets. It is notable that the highest overall performance was reported by Hu et al. [49], with an AUC of 99%, using one of the largest datasets, with 11,118 image frames, as compared to the other reviewed studies.

Polyp and tumour detection methods

Polyps are growing protrusions of mucosa inside the intestine due to excessive proliferation of tissue and inflammation or deep-seated malformations. Polyps are mainly discriminated by their shape and texture [49]. Mamonov et al. [51] started by accepting polyps as protrusions that are mostly round in shape. Thus, the best-fit ball radius was used as a decision parameter of a classifier. Yuan et al. [52] use the Scale-Invariant Feature Transform (SIFT) to detect salient points in images that may correspond to polyps. From the neighbourhoods of these points, texture features, called Complete Local Binary Pattern (CLBP), are extracted. A more complex methodology was proposed by Liu et al. [53] for small-bowel tumour detection. This methodology is based on both texture and colour analysis. Multi-scale texture analysis is performed (i.e. the analysis is performed at different image resolutions) by means of the curvelet transformation and fractal encoding. Colour information is captured by means of higher order moments between different colour channels. The extracted features are classified by an optimally selected SVM classifier. In another study, Alizadeh et al. [54] proposed an adaptive neuro-fuzzy inference system aimed at classifying CE video frames containing polyps. The system extracts 32 features including four statistical measures (namely, contrast, correlation, homogeneity, and energy) calculated from co-occurrence matrices. Mutual information (a measure of the information shared between pairs of features) was used to select a subset of more informative features for the discrimination of polyps from normal tissues.

The results of the polyp/tumour detection methods reviewed in this subsection are presented in Figure 1. The highest results were reported by Liu et al. [53], who achieved an accuracy of 97.3%. However, the dataset used is smaller than the one used by Hu et al. [49], who achieved an accuracy of 92.4%. The lower accuracy of the latter can be mainly attributed to its lower specificity.

Figure 1

Graphical representation and comparison of the aforementioned detection methods for the different kind of abnormalities during the last 5 years (2014–2018). The vertical bars indicate overall ranges of accuracy, sensitivity and specificity obtained from various abnormality detection methods. Single points indicate no variance. Missing bars indicate unavailable data due to inconsistent reporting of results across studies

Ulcer detection methods

Ulcers are among the most common pathologic outcomes of several diseases affecting the GI tract. Small-bowel ulcers, of variable severity and activity, are often challenging in terms of detection by traditional imaging techniques. Hence, CE is increasingly being used in ulcer diagnosis and management. Several solutions have been proposed for the detection and follow-up of different forms of CD or NSAID-induced ulcerations. In this context, the majority of recent studies were based on combinations of texture and colour features to discriminate ulcers from normal tissues. In the study of Suman et al. [55], the colour components of seven different colour spaces were analysed in order to find an optimal combination of their colour components that better discriminates ulcers from normal tissues. The selected components were Cr (which represents the difference of red from a reference value) from YCbCr colour space, the yellow (Y) component from CMYK colour space, and the blue (B) component from RGB. A two-stage, fully-automated, computer-aided detection system is proposed by Yuan et al. [56] to detect ulcers in CE images. In the first stage, the image is segmented using superpixels of different sizes. Colour and texture features are extracted from each superpixel. Then, the extracted features are fused to form a saliency map per image. In the second stage the BoW technique uses the obtained saliency map to better characterise the images. A system capable of using weakly annotated images was proposed by Vasilakakis et al. [57]. Instead of annotating the images in detail, i.e. pixel-by-pixel, the images can be annotated at image level. In this way, a binary semantic label is assigned per image to indicate whether its content is normal or abnormal, e.g. the keyword “abnormal” if the image contains an abnormality. This system offers the convenience to robustly detect which images contain possible ulcers. A weak image annotation method based on a CNN architecture was presented by Georgakopoulos et al. [58]. The CNN receives a single CE image as input to process, and it makes the decision regarding the presence of an inflammatory lesion in the input image (i.e. two-neuron output layer).

Figure 1 gathers the results from the previously presented studies, which focus on the ulcer detection. Suman et al. [55] reported the best results compared to the other methods, achieving an accuracy of 97.9%. However, the results cannot be directly compared to the other methods because different datasets were used. In the method of Georgakopoulos et al. [58], where a CNN architecture is utilised, the performance is noticeably high, with an accuracy of 90.2%. Although the dataset in the study of Georgakopoulos et al. [58] consists of weakly annotated images, automatically extracted features of the proposed method are able to characterise the images and provide a classification system independent of handcrafted features and pixel-level image annotation.

Hookworm detection methods

Hookworms are a leading cause of maternal and child morbidity, and their detection is often a challenging task. An automatic hookworm detection system for CE images was proposed by Wu et al. [59]. The system is based on the basic characteristics of hookworms, which are their tubular body structure, the parallel edges in the shape of their body, and their body colour features. The first step is the processing of the image for the enhancement and the detection of locations with tubular structure. The second step is a parallel region detection method identifying the potential regions having hookworm bodies. The last step for the detection of hookworms is the features based on their intensity, which discriminate the hookworms from different components of the GI tract, i.e. bubbles. A histogram of average intensity is proposed to represent their properties. A deep hookworm detection system based on CNN was proposed by He et al. [60] for CE images. The system consists of two CNN networks, which enable the simultaneous modelling of visual appearances and tubular patterns of hookworms.

The results of the proposed methods are shown in Figure 1. The experiments were performed in the same dataset for all the aforementioned hookworm detection methods. It can been noticed that the method of He et al. [60] for the hookworm detection achieved higher classification performance than the method of Wu et al. [59]. This can be attributed to the better enhancement of the visual pattern of the hookworms achieved by the CNN.

Multiple lesion detection methods

The detection of different lesions is one of the most challenging tasks in the CE reviewing process. To this end, the majority of research studies focus on the detection of one kind of abnormality, probably because it is easier for an algorithm to distinguish one abnormality each time, i.e. bleeding detection is mainly based on the characteristic red hue of blood. Nawarathna et al. proposed [61] a texture analysis method. Different texture features, i.e. LBP, of an image extracted and the distribution of these various texture features is captured by a histogram to characterise the content of each image. An automated technique was proposed [62] for abnormality detection in CE images. Every image is divided into several blocks and from each block the colour information is extracted using a CNN to overcome the drawbacks of handcrafted features. In [63], the SIFT features are extracted from images in HSV colour space to obtain visual words that represent the three lesion categories (bleeding, polyps, ulcer) and the normal images. Then, these four types of visual words are combined together to composite the representative visual words for classifying the CE images. It is known that the semantics of the normal content include mucosal tissues, the hole of the lumen, bubbles, and debris. Thus, an investigation of such a semantic interpretation of the CE images may yield an improvement in lesion detection. In [64], a system for the semantic interpretation of the whole CE content was proposed, in which each semantic content category, i.e. the hole of the lumen it represents a different label. The system consists of a salient point detection algorithm to detect points of interest in weakly annotated images, and it extracts colour features from the area around them. The representation of the CE images is based on the BoW image representation technique. Then, multi-label SVM classifiers are utilised to discriminate the labels that exist in a CE image. The study of [64] was further extended in [65], where a convolutional neural network architecture enabling multi-scale feature extraction (MM-CNN) was proposed to detect the existing labels in the CE images. In [66], a three-phase methodology based on a CNN architecture for automatic detection and localisation of lesions in CE images was presented. In the first phase, the proposed CNN architecture automatically extracts features from weakly annotated images. In the second phase, it suggests the possible locations of lesions in the detected CE images. In the third phase, a new algorithm is proposed to localise the lesions in the detected CE images. This algorithm uses the automatically extracted features from CNN. Then, the algorithm detects the suggested locations from the second phase, which belong to lesions.

A summary of previously presented methods for the detection of different kinds abnormalities is in Figure 1. It can be noticed that the method proposed by Vasilakakis et al. [65] has considerably higher results, with an AUC of 90.0%, which indicates that the semantic interpretation of the content of CE images can provide valuable information to assist in the detection of different lesions.