Correlation of Endoscopic Ultrasound and Texture Analysis of Endoscopic Images in Esophageal Cancer Staging

Introduction: EUS has been shown to be the most accurate technique to determine tumor depth of infiltration. However, understaging and overstaging can occur in EUS staging. Texture analysis has been applied in ultrasonic tissue characterization of liver. We tried to evaluate the feasibility of using texture analysis of endoscopic imaging in aiding cancer staging in addition to EUS.

Aims and Methods: Endoscopic images of esophageal cancer were obtained to evaluate whether the T stage can be determined by texture analysis of tumor images. From Oct. 2001 to Jan. 2006, 32 patients (28 male and 4 female, mean age 61.5 years) with esophageal cancer underwent EUS staging were included. Three to five 64 X 64 pixels endoscopic sub-images were selected at tumor area from original images in each patient. Total of 145 sub-images were derived. Image features of the green component of the bitmap sub-images were extracted by gray level concurrence and non-separable wavelet transform. The tumor stage was determined by endoscopic ultrasound using UM DP12–25R, UM2R, or MH908 (Olympus Optical Co.)

Results: Twelve cases were staging as T1 by EUS, one was T2, and nineteen were T3. The tumor stage was divided into two classes: early (T1) and advanced (T2 and above). Classification using the extracted image features by a classifier known as the support vector machine was tested with leave-one-out method and was correlated with the T staging of EUS. The best classification accuracy was 75.2% (63.3% for early class and 83.5% for advanced class) by endoscopic image texture analysis.

Conclusion: For esophageal cancer T staging, EUS is superior to endoscopic image texture analysis.