CVD_Net: Head and Neck Tumor Segmentation and Generalization in PET/CT Scans Across Data from Multiple Medical Centers

Accurate diagnosis, analysis, and monitoring of the progress of head and neck squamous cell carcinoma (HNSCC) or tumors using Positron Emission Tomography (PET) and Computed Tomography (CT) is paramount for radiation therapy treatment plans. Deep learning methods have shown promising performances in segmenting HNSCC, but most of the methods are trained and tested on homogeneous datasets (data from the same source), leading to a degradation of performance when tested on independent datasets from different sources, as in real-world scenarios. One of the main causes of the poor generalizability is the variability in the quality of scans from diverse sources. In this work, we propose a novel algorithm, CVD_Net, which is a combination of Convolutional Neural Networks for feature extraction, Vision Transformers to capture long-range dependencies, and domain-specific adapters to mitigate the problem of negative knowledge transfer. The CVD_Net was evaluated on the HECKTOR 2022 dataset collected from nine medical centers around the world, obtaining a mean Dice Score of 0.77492 (comparable to performances from specifically designed state-of-the-art algorithms) when tested on a hidden dataset from three medical centers, two of which are new (not seen during training). Furthermore, the algorithm also demonstrated high generalizability performance when tested on independent data from a new medical center.