Classification of esophageal cancer by using hyperspectral data.

Purpose: Esophageal cancer is widespread worldwide, with the highest rate in Asia. Early diagnosis plays a key role in increasing the survival rate. Early cancer detection as well as fast evaluation of tumor extent before and resection margins during/after surgery are important to improve patients' outcomes. Hyperspectral imaging (HSI), as a noninvasive and contactless novel intraoperative technique, has shown promising results in cancer detecting in combination with artificial intelligence.

Methods: In this clinical study, the extent to which physiological parameters, such as water or hemoglobin content, differ in the esophagus, stomach, and cancer tissue, was examined. For this purpose, hyperspectral intraluminal recordings of affected tissue specimen were carried out. In addition, a classification of the three intraluminal tissue types (esophageal, stomach mucosa, and cancerous tissue) was performed by using two different convolutional neural networks.

Results: Our analysis clearly demonstrated differences in hemoglobin concentration and water content between healthy and cancerous tissues, as well as among different tumor stages. As classification results, an averaged area under the curve score of 81 ± 3%, a sensitivity of 74 ± 8%, and a specificity of 89 ± 2% could be achieved across all tissue types using a hybrid convolutional neural network.

Conclusion: HSI has relevant potential for supporting the detection of tumorous tissue in esophageal cancer. However, further analyses including more detailed histopathologic correlation as "gold standard" are needed. Data augmentation and future multicenter studies have to be carried out. These steps may help to improve and sharpen our current findings, especially for esophageal cancerous tissue.