Computer-aided detection of early neoplasia in the esophagus using high definition endoscopic images

INTRODUCTION: adenocarcinoma of the esophagus is the fastest rising type of cancer in the Western world. The recent development of High Definition (HD) endoscopy has enabled the specialist physician to identify Barrett's cancer at an early stage. Nevertheless, it still requires considerable effort, training and expertise to be able to recognize these irregularities associated with early cancer. AIMS & METHODS: As a first step towards a Computer-Aided Detection (CAD) system that aids the physician in finding these early stages of cancer, we constructed an algorithm that is able to identify irregularities in the esophagus automatically, using HD endoscopic images. The concept employs tile-based processing, and the system is not only able to identify that an endoscopic image contains early cancer, but it can also delineate the abnormality. The detection uses image information on color as well as texture (surface irregularity).The identification is based on the following steps: (1) pre-processing, (2) feature extraction with dimensionality reduction, (3) classification. We evaluate the performance in different color spaces using the Color Histogram (CH) and Gabor features for the detection of early stages of cancer. For classification, we employ a Support Vector Machine (SVM) and evaluate its performance using different parameters and kernel functions. RESULTS: For clinical evaluation of the proposed algorithm, we gathered endoscopic images of 66 patients and let an expert physician delineate the irregularities. After splitting the images into tiles, we selected 600 tiles for each class, i.e. tumorous and normal tissue for every tile size. Using these tiles, we trained a classifier for the detection for the detection of tumorous tissue. We used 40% of the total data set for training and the other 60% for testing. Table 1 shows the classification performance of our classifier on the test sets. For the tile sizes larger than 75 × 75 pixels, the performance dropped significantly. For this reason, these results are omitted. The system achieves a classification accuracy 95.9% on 50x50 pixel tiles of normal and of tumorous tissue and reaches an Area Under the Curve (AUC) of 0.990.