Abstract
Background We developed a computer-assisted diagnosis model to evaluate the feasibility of automated
classification of intrapapillary capillary loops (IPCLs) to improve the detection
of esophageal squamous cell carcinoma (ESCC).
Methods We recruited patients who underwent magnifying endoscopy with narrow-band imaging
for evaluation of a suspicious esophageal condition. Case images were evaluated to
establish a gold standard IPCL classification according to the endoscopic diagnosis
and histological findings. A double-labeling fully convolutional network (FCN) was
developed for image segmentation. Diagnostic performance of the model was compared
with that of endoscopists grouped according to years of experience (senior > 15 years;
mid level 10 – 15 years; junior 5 – 10 years).
Results Of the 1383 lesions in the study, the mean accuracies of IPCL classification were
92.0 %, 82.0 %, and 73.3 %, for the senior, mid level, and junior groups, respectively.
The mean diagnostic accuracy of the model was 89.2 % and 93.0 % at the lesion and
pixel levels, respectively. The interobserver agreement between the model and the
gold standard was substantial (kappa value, 0.719). The accuracy of the model for
inflammatory lesions (92.5 %) was superior to that of the mid level (88.1 %) and junior
(86.3 %) groups (P < 0.001). For malignant lesions, the accuracy of the model (B1, 87.6 %; B2, 93.9 %)
was significantly higher than that of the mid level (B1, 79.1 %; B2, 90.0 %) and junior
(B1, 69.2 %; B2, 79.3 %) groups (P < 0.001).
Conclusions Double-labeling FCN automated IPCL recognition was feasible and could facilitate
early detection of ESCC.
