Mo1671 a decision-tree classifier to improve the accuracy of magnification endoscopic assessment of lateral spreading tumours

Takamaru, Hiroyuki, Stammers, Matthew, Sekiguchi, Masau, Yamada, Masayoshi, Sakamoto, Taku, Matsuda, Takahisa and Saito, Yutaka (2019) Mo1671 a decision-tree classifier to improve the accuracy of magnification endoscopic assessment of lateral spreading tumours. Gastrointestinal Endoscopy, 89 (6), [AB514]. (doi:10.1016/j.gie.2019.03.861).

Abstract

Introduction: it is well known that when performing endoscopy for depth diagnosis of T1 colorectal cancer (CRC) one sometimes experiences, the difficulty of depth diagnosis, even with magnification. This is especially the case in lesions containing large nodules. It has been reported that the location of submucosal invasion in laterally spreading tumors (LST) varies between granular type (G) and non-granular type (NG).

We hypothesize that the contribution of pit pattern diagnosis and JNET classification using magnified endoscopy is different between LST-G and LST-NG.

Methods: a total of 647 LSTs in 612 patients with diagnosed tumor in-situ (Tis) or T1 diagnosed by magnified endoscopy using both pit-pattern and JNET classification were analyzed retrospectively.

All lesions were either endoscopically or surgically resected at our institution between Jan 2015 and Dec 2017. All endoscopic findings were recorded in the “Japan Endoscopic Database (JED)”. Independent variables included: JNET classification, macroscopic features, endoscopically estimated lesion size for lesions T1b or deeper. The lesions were divided into LST-G or LST-NG. Histological diagnosis was used as the gold standard for assessing the depth of invasion.

The lesions were randomly split 50-50 into test and training datasets and a decision tree classifier was trained on each group using the training data. Then the model was deployed on the test set and a receiver operator curve (ROC) was calculated for each model’s performance on the test set.

Results: among all the LSTs, mean size of the lesions were 27.5mm. The ratio of macroscopic features was as follows;

The number of LST-G’s was 369 and LST-NG’s 278.

Lesions of T1b or deeper were included 91 (24.7%) and 76 (37.3%).

The AUC of ROC for LST-G’s was 0.892 in the training data set and 0.846 in the test data set. The weighted variable contribution to the algorithm for the diagnosis of T1b or deeper was as follows; Pit pattern: 0.74, JNET: 0.19, macroscopic features: 0.08, and size of the lesion: 0.0.

On the other hand, The AUC for LST-NG’s was 0.931 in the training data set and 0.938 in the validation data set. The variable contribution was as follows; Pit pattern: 0.92, JNET: 0.05, size: 0.02 and macroscopic features: 0.0.

The decision tree of LST-NG, a combination of endoscopic findings showed 73% to 96% sensitivity for T1b or deeper and 84% to 98% specificity.

LST-G demonstrated 86% sensitivity and 54% to 95% specificity. The specificity was lowest for 0-Is or 0-Is+IIc lesions.

Conclusion: pit patterns contributed to the diagnosis of T1b or deeper in both LST-G and LST-NG models. In the case of LST-G with 0-Is component, it appears that depth diagnosis difficult regardless of the size of the lesions. Further research is warranted in this area in order to improve things further.

This record has no associated files available for download.

Contributors

Download statistics