An Architecture and a New Deep Learning Method for Head and Neck Cancer Prognosis by Analyzing Serial Positron Emission Tomography Images

In the U.S. it is estimated that there are more than 20,000 cases of head and neck cancers per year. Radiomics is a much discussed topic in nuclear medicine. The radiomic characteristics of metabolic imaging modalities such as Positron Emission Tomography (PET) have been postulated as surrogates for underlying tumor biology and thus prognosis. Radiomic data can be extracted to discover characteristics and patterns of evolution (in serial images, their changes over time) and to provide a response to treatment. In oncology it has been shown that the degree of tumor heterogeneity is a prognostic factor for survival and an obstacle to cancer control. One of the main obstacles to radiomics research is the lack of understanding among clinicians and data scientists. For this reason, in this paper, we propose a case study, an architecture and a Deep Learning method for the processing and analysis of PET tomographic images for the detection of head and neck cancers. Our architecture consists of three phases: 1) Image preparation, 2) Deep learning method using convolutional neural networks for dimensionality reduction and image feature extraction, and recurrent neural networks for serial image learning of PET, and 3) Optimization. A public dataset is used and the quality of the method is demonstrated using standard quality measures such as Accuracy, Precision, Recall and F1-Score.