ADTH-07 Deep learning based classification of intrapapillary capillary loops for detection of early oesophageal squamous neoplasia

Abstract

Introduction Narrow band imaging with magnification endoscopy (ME-NBI) allows detailed assessment of microvascular patterns within early squamous cell neoplasia (ESCN). The Japanese Endoscopy Society (JES) AB classification describes ESCN and predicts invasion depth based on the intrapapillary capillary loop (IPCL) pattern. Early lesions are amenable to endoscopic therapy (EET). We have designed a novel deep convolutional network (CNN) able to classify IPCL patterns as normal (A) or abnormal (B1/B2/B3), in order to alert endoscopists to abnormal areas in ESCN during lesion assessment. Methods 17 patients were recruited at a referral centre in Taiwan. Endoscopies were performed using ME-NBI (Olympus). IPCLs were classified for each video by 2 experts as normal (type A), or abnormal (type B1/B2/B3), using the JES classification. Matched tissue was obtained by ESD for histologic evaluation of invasion depth. Images were quality controlled to remove uninformative (blurred) images. Our full dataset consisted of 7046 images. A CNN was developed with five-fold cross validation. On average, each fold used 3962 images for training, and 1637 unseen images (846 normal and 791 abnormal) for testing. Accuracy, F1 scores, sensitivity and specificity for abnormal IPCL detection were calculated. Results 17 patients were included (10 had early neoplasia [1 high grade intra-epithelial neoplasia (HGIN) 4 carcinoma in situ (CIS) invading to the lamina propria, 4 to the muscularis mucosa and 1 to the submucosa] and 7 were normal). Our algorithm operates at video rate and had an accuracy for differentiating abnormal IPCL patterns (type B1, B2, B3) from normal (type A) of 93.69%. The average F1 score for identifying abnormal areas of ESCN based on IPCL classification was similarly high at 92.2%. Our network also achieved a sensitivity and specificity for abnormal IPCL detection of 89.3% and 98% respectively. Conclusion We introduce a novel application of deep learning by developing a real-time CNN, with promising results in classifying squamous mucosa as normal or neoplastic based on the JES IPCL classification. Our system demonstrates impressive accuracy, sensitivity and specificity for differentiating type A from type B1/2/3 IPCLs. Further work is underway to develop a multiclass classifier to distinguish between the subtypes of IPCL patterns. Such a validated system could be used in vivo to alert endoscopists to the presence of ESCN and direct planning of appropriate EET.