Software Miroview 4.5 Express - Validação

Abstract: BACKGROUND: Reading capsule endoscopy (CE) is time-consuming. The Express View (EV) (IntroMedic, Seoul, Korea) software was designed to shorten CE video reading. Our primary aim was to evaluate the diagnostic accuracy of EV in detecting significant small-bowel (SB) lesions. We also compared the reading times with EV mode and standard reading (SR).

METHODS: 126 patients
with suspected SB bleeding and/or suspected neoplasia were prospectively enrolled and underwent SB CE (MiroCam®1200, IntroMedic, Seoul, Korea). CE evaluation was performed in standard and EV mode. In case of discrepancies between SR and EV readings, a consensus was reached after reviewing the video segments and the findings were re-classified. RESULTS: The completion rate of SB CE in our cohort was 86.5% and no retention occurred. The per-patient analysis of sensitivity, specificity, positive predictive value, negative predictive value, and diagnostic accuracy of EV compared to SR were 86%, 86%, 90%, 81%, and 86%, respectively, before consensus. After consensus, they increased to 97%, 100%, 100%, 96%, and 98%, respectively. The median reading time with SR and EV was 71 min (range 26- 340) and 13 min (range 3-85), respectively (p < 0.001). CONCLUSIONS: The new-generation EV shows high diagnostic accuracy and significantly reduces CE reading times.

1. Introduction

Capsule endoscopy (CE) is a well-established method to explore the small bowel (SB).
National and International guidelines recommend performing CE in case of suspected SB
bleeding (SSBB), suspected or known Crohns disease, suspected SB tumors, inherited
polyposis syndromes, and celiac disease [1,2]. Although continuous technological advances
during the last three years improved CE image quality and battery life, CE reading remains
time-consuming, with mean reading times of 50-75 min (ranging from 30 to 120 min) and
based on the readers experience [3-9].

Reading requires the clinicians to pay and maintain attention for a long time without distractions [10]. A recent study showed that the fatigue of experienced readers significantly affects their accuracy, causing a decrease of 20% in lesion detection after the first CE study [11].

To shorten reading times, CE readers often increase the frame rate. However, this strategy inevitably leads to an increased and unacceptable miss rate by reducing significant lesion detection [12]. Innovative algorithms have been developed to shorten video evaluation by removing redundant images based upon their similarity, thus reducing reading time, such as RAPID®, QuickView (PillCam, Medtronic, Dublin, Ireland) [13-18], Omni Mode (Endocapsule, Olympus, Tokyo, Japan) [5,6], MiroViewTM, and Express View (MiroCam®, IntroMedic, Seoul, Korea) [8,19].

Dedicated gastroenterologists have accepted these technologies, as shown in a recent international survey by the I CARE (CApsule endoscopy Research) group. However, they also pointed out the need for CE readers to implement artificial intelligence (AI) systems further [20].

Express View (EV) is a reading software designed to reduce CE video by simultaneously applying a function that recognizes abnormal findings (suspected disease detection) and another function that enables the cut of insignificant redundant images (video summary). This software was initially validated by Saurin et al. [19] and by Gomes et al. [8] on 83 and 89 patients, respectively, showing 82.2-83.1% sensitivity for detection of clinically significant lesions.

Recently, a third-generation artificial intelligence-based EV algorithm, working with a binary map and a computational 8-ray connectivity, was released. The abnormal detection function, based on a conventional machine learning (ML) algorithm that uses a feature descriptor set of color, texture, and shape, was reinforced and tunned with other lesions (bleeding, ulcer, vascular, polyp) to implement local features for ROI (region of interest). Moreover, the video summary function, which can convert the red-green-blue (RGB) color model into the hue-saturation-value (HSV) to further compare images and skip similar ones, was improved.

Implementing the histogram comparison of the local binary pattern (LBP) format, the video summary allows the removal of meaningless images, providing a compression rate of the initial video of up to 95% (compared to 80% of EV version 2.0). This multicenter study aims to prospectively evaluate the diagnostic accuracy of the new-generation EV in a per-patient and per-lesion analysis and to evaluate its impact on reading time by comparing the time of reading in EV mode and standard mode.

2. Materials and Methods

2.1. Population and Procedure

Patients undergoing CE in 6 Italian centers (Brescia, Ravenna, Feltre, Aviano, Torino, Ferrara) were enrolled for this study between January 2019 and July 2020. Each center consecutively enrolled 20 patients, except for the coordinating center (Brescia), which consecutively enrolled 26 patients. All patients signed the required informed consent form.

Indications for CE were SSBB (overt or occult) with negative bi-directional endoscopy and/or suspected SB neoplasia (based on symptoms and previous radiological tests) or polyposis. CE was performed with the MiroCam®1200 (IntroMedic, Seoul, Korea), which provides a 320 × 320 resolution, a field of view of 170°, and an image acquisition rate of 3 frames per second (fps). Each participating center was allowed to follow its preprocedural preparation protocol if based on the current international guidelines [1,21].

Therefore, all patients had 2 L polyethylene-glycol (PEG)-based bowel prep as a split. regimen, simethicone 30 min before capsule ingestion, and prokinetics administered orally in case the capsule was found in the stomach, 1 h after its ingestion, with real-time viewer check.

CE was considered complete if the capsule reached the caecum within the recording time. All patients with a complete SB transit and those in whom the capsule entered the SB (even if not reaching the caecum) were included in the analysis. In addition, capsule retention data, defined as a capsule that was not excreted within 15 days, and complication rates, were collected.

2.2. Capsule Reading

Each video was initially evaluated in standard view by one experienced (>200 cases) local reader from the enrolling center (standard reading, SR). A single frame reading at 20- 25 fps speed was used for stomach review, whereas the SB reading was performed at a maximum speed of 10-12 fps, as suggested by the ESGE Guidelines [21,22]. Once  generated, anonymized videos were sent to the "coupled center," chosen among the other centers, for the following blinded reading in EV mode (EV Version 3 2018).

The shortened continuous video (removed by redundant frames) processed by the EV software was reviewed by the blinded reader with the same modalities of SR. For both standard and EV reading, the readers collected clinical information and findings in a case report form (CRF); the bowel cleansing was also reported using the QI (Quantitative Index) Brotz scale [23].

Positive findings were classified in terms of type, location, and relevance as P0-P2 (according to the Saurin classification) [24]. Reading time was defined as the time (in min) spent watching the video (in both modalities), including the time for selection of images, without considering the time spent to write the final report. Reading time for the SB examination was recorded and reported separately from the esophagus and gastric reading time.

2.3. Interpretation of the Results

A consensus meeting of designated experts among the principal investigators of the involved centers (AM, CS, and SP, each > 400 CE reading experience) examined all cases by comparing the CRFs filled by the standard reader with those filled by the blinded reader in EV mode. In the case of multiple findings corresponding to different diagnoses and  clinical relevance, the most significant lesion supporting the reported diagnosis was specified.

Concordant cases for positive or negative findings (final diagnosis and/or detected lesions)
were not re-evaluated. Conversely, in case of discrepancies between the two reading modes, a panel of experts re-evaluated the videos and adjudicated. Specifically, a face-to-face meeting was organized, and the 3 consensus panel members simultaneously re-evaluated the full videos of the discordant cases using both the standard and EV reading, and eventually reclassified the findings.

The primary endpoint was to evaluate EV sensitivity in detecting SB lesions (P1 and P2, according to the Saurin classification) before and after consensus review with per-patient and per-lesion analysis. SR was first considered the reference standard and used as a comparison to measure the reading agreement. After consensus, EV accuracy was re-evaluated considering SR after consensus review the new reference standard.

Secondary endpoints were (1) median reading time with EV and SR, (2) evaluation of EV accuracy in detecting significant SB findings (P1 and P2 lesions) before and after  consensus, and (3) diagnostic yield (DY) of EV and SR when compared to consensus review (Figure 1).

2.4. Statistical Analysis

Reading time for EV and SR was expressed as a median (min and max value range).
Means and standard deviations were used to describe continuous variables. Proportions
were used to describe categorical variables. A Mann-Whitney test was used to compare
the median reading time in EV mode and SR. EV accuracy was measured in terms of
sensitivity (sens), specificity (spec), positive predictive value (PPV), negative predictive
value (NPV), and diagnostic accuracy, all presented as percentages (95% confidence
interval, CI) at the per-patient and per-lesion analysis.

The final diagnosis was considered for the per-patient analysis. In the case of a CE video showing numerous (>10) repetitive SB lesions (including Crohns disease with numerous ulcerations, or diffuse angioectasias), these findings were considered as corresponding to one single lesion for the per-lesion analysis. Although variability exists in the literature, meta-analyses and larger studies suggest about a 60% DY for SB CE in patients with SSBB [2].

Therefore, by assuming a 60% prevalence of clinically relevant lesions, a sample size of 121 patients would estimate sensitivity of 90% of the new method with a 95% CI between 83% and 97%. Considering a failure rate of 5%, the total sample size was 126 patients.

3. Results

In the present study 126 patients were enrolled (mean age 67.6 ± 14.6; 64 women).
Indications for CE were occult SSBB in 99 patients (78.6%), overt SSBB in 25 patients
(19.8%), and suspected SB neoplasia in 2 patients (1.6%). Technical issues (i.e., recording
was temporarily interrupted during capsule transit in the stomach) were reported in one;
CE was complete in 109 out of 126 patients (86.5%). In 17 patients the capsule did not reach the caecum.

Incomplete examination occurred in 11 out of 17 patients due to delayed SB transit and in 2 patients due to SB strictures (n = 1 duodenal fibrotic ulcer; n = 1 jejunal adenocarcinoma). In the remaining four cases, the incomplete examination was related to delayed gastric transit (last CE image was acquired in the stomach). No case of capsule retention occurred. In one case, the capsule was endoscopically retrieved after 4 days.

No adverse events were reported. Mean gastric and SB transit times were 40.4 min (range 0-264) and 293.3 min (range 23-690), respectively (Table 1). A total of 98 out of 122 patients with SB evaluation had a QI Brotz scale = 6. Mean QI Brotz scores given by readers in standard mode and EV mode were 7.11 ± 1.93 and 6.96 ± 1.86, respectively.

3.1. Per-Patient Analysis

One hundred twenty-two (96.8%) CEs were included in the analysis. Before consensus,
SR and EV agreed with the final diagnosis in 105 cases (n = 62 positive for SB finding and n= 43 negative). The remaining 17 cases had discordant results. In particular, 10 had a positive SR (SR+) and negative EV reading (EV-); 7 had a negative SR (SR-) and positive EV reading (EV+). Before consensus, EV sensitivity, specificity, PPV, and NPV were 86%, 86%, 90%, and 81%, respectively. The overall EV diagnostic accuracy was 86% (Table 2). After consensus,the 17 discordant cases were reclassified as following: All seven cases SR-EV+ were reclassified as positive at both SR and EV mode (SR+EV+); of the remaining 10 SR+EV- cases, seven were reclassified SR+EV+ and one was reclassified as SR-EV-; and the remaining two cases were confirmed to have discordant reading (SR+EV-) since the EV software inappropriately cut the findings responsible for the main diagnosis (an ileal hemangioma and an ileal erosion). As a result, 78 patients were reclassified as having a final diagnosis affecting the SB, whereas 44 patients had no significant pathological SB findings (Table 3 and Figure 2). After consensus, EV sensitivity, specificity, PPV, and NPV were 97%, 100%, 100%, and 96%, respectively, with an overall diagnostic accuracy of 98% (Table 2). The accuracy of EV and SR after consensus (representing the reference standard) are reported in Table 4.