Detection of Early Esophageal Cancer from Endoscopic Images Based on a Haar Wavelet Feature

We propose a new method for early esophageal cancer detection from endoscopic images. In the proposed method, an endoscopic image is converted to the CIE $\mathrm {L}^{*}\mathrm {a}^{*}\mathrm {b}^{*}$ color space, and the Haar wavelet transform is applied to the $\mathrm {L}^{*}$ and $\mathrm {a}^{*}$ components. First, we create an average image of the normal region from the $\mathrm {a}^{*}$ component. Next, we calculate the threshold for detecting abnormal regions from the average image, based on a box plot. In our experiment, the $\mathrm {L}^{*}$ and $\mathrm {a}^{*}$ components of the endoscopic image are divided into small blocks. The $\mathrm {L}^{*}$ component is normalized and binarized, to determine the analysis target. The a*component is used to calculate a trim mean, and this is compared with a threshold and binarized. Then, the logical product of the $\mathrm {L}^{*}$ and $\mathrm {a}^{*}$ components is computed to generate an enhanced image and detect abnormal regions. We describe the method for detecting abnormal regions in detail, and show that our proposed method is useful for early esophageal cancer detection from endoscopic images.