Computer-aided detection (CADe) and diagnosis (CADx) systems such as CAD EYE (EX-1, Fujifilm), launched in 2020, have been widely adopted worldwide and demonstrated significant value in colorectal neoplasia detection and characterization 1,2). In 2024, a new endoscopy system, EP-8000 (Fujifilm), became available, providing brighter images with reduced halation and improved contrast 3). In the same year, a new CAD EYE software version with a 1.6-fold larger training dataset and improved performance was released. This study aimed to evaluate the detection and diagnostic performance of CAD EYE with the EP-8000 system and compare the results with those obtained using the previous system (VP-7000) and the earlier software version.

This single-center observational study included consecutive colorectal lesions (2–10 mm) detected by expert endoscopists using EP-8000 from March to May 2025. After lesion recognition by the endoscopist, CAD EYE was activated. White-light imaging (WLI) and linked color imaging (LCI) with and without 1.5× high-speed scanning (high-speed WLI/LCI), including approximately 3 cm of surrounding mucosa, were evaluated for CADe performance. Positive detection was defined as the accurate appearance of an annotation box on the lesion. CADx performance was evaluated using CAD EYE-assisted BLI-magnifying observation and compared with subsequent histopathology. Outcomes were compared with 100 previously evaluated lesions obtained under VP-7000 using the same methodology.

A total of 27 patients with 100 lesions were analyzed (mean age 66.7±10.9 years; males 66.7%). Mean tumor size was 4.2±2.7 mm; proximal colon 46.0%; protruded type 41.0%. Histology included sessile serrated lesions and low-grade adenomas: 57 and 43 lesions, respectively. Detection rates were: WLI 94.0% vs. high-speed WLI 68.0% (p<0.001), LCI 94.0% vs. high-speed LCI 75.0% (p<0.001), WLI 94.0% vs. LCI 94.0% (p=1.000), and high-speed WLI 68.0% vs. high-speed LCI 75.0% (p=0.273).CADx accuracy was 95.0%, slightly lower than expert BLI-magnifying diagnosis (100.0%, p=0.06). Comparisons with VP-7000 showed no differences in lesion characteristics, while CADe was significantly improved under EP-8000 in WLI: 94.0% vs. 85.0% (p=0.038). Other detection modes showed comparable performance. CADx accuracy tended to improve: 95.0% vs. 87.8% (p=0.080). Importantly, false-positive alarms per case were significantly reduced with EP-8000: 12.4±7.9 vs. 19.5±13.4 (p=0.020).

CAD EYE with the new EP-8000 system demonstrated improved detection and diagnostic performance compared to the previous system, along with fewer false alarms. Further data accumulation is ongoing to strengthen these findings.