Abstract:
CNNs have immense potential for transforming conventional image processing methods in
medical imaging. They are gaining popularity among radiologists and physicians for specializing
in processing image data but the development of such intelligent systems for use in the
healthcare industry in general, is bottlenecked due to challenges that include limited data
availability and data privacy laws like Health Insurance Portability and Accountability Act
(HIPAA) in the US. We propose the use of self-configuring pipelines like nnUNet which can
empower healthcare professionals like physicians and technologists to train and deploy models
locally, without expert domain knowledge in Machine Learning and with no need to share
confidential patient data. This research work exploits and evaluates a self-optimizing deep
learning framework, nnUNet for the automated semantic segmentation of tumors and lymph
nodes on 3D FDG PET/CT images of Head and Neck. nnUNet is an adaptive segmentation
method that examines the provided training examples and customizes a UNet-based
segmentation pipeline, tailored to the requirements of the given dataset. Since manual delineation
of cancerous lesions is time consuming and error prone, automated segmentation can flag
concerning areas on the images, allowing radiation oncologists to optimize the treatment plan in
a shorter time and with increased reliability. Our method achieved an aggregated dice score of
0.77 for GTVp and 0.67 for GTVn (0.72 on average) using nnUNet, which unveiled comparable
performance to the best performing models for Head and Neck tumor segmentation. By
automating common tasks like data preprocessing, network architecture selection and
hyperparameter optimization, self-configuring methods can enhance the adoption of AI in the
healthcare industry by making healthcare professionals embrace the latest AI technologies,
bridging the gap between AI industry and healthcare.
