Skip to content

2Be3-Net : Combining 2D and 3D convolutional neural networks for 3D PET scans predictions

Radiomics – high-dimensional features extracted from clinical images – is the main approach used to develop predictive models based on 3D Positron Emission Tomography (PET) scans of patients suffering from cancer. Radiomics extraction relies on an accurate segmentation of the tumoral region, which is a time-consuming task subject to inter-observer variability. On the other hand, data-driven approaches such as deep convolutional neural networks (CNN) struggle to achieve great performances on PET images due to the absence of available large PET datasets combined to the size of 3D networks. In this paper, we assemble several public datasets to create a PET dataset large of 2800 scans and propose a deep learning architecture named “2Be3-Net” associating a 2D feature extractor to a 3D CNN predictor. First, we take advantage of a 2D pre-trained model to extract feature maps out of 2D PET slices. Then we apply a 3D CNN on top of the concatenation of the previously extracted feature maps to compute patient-wise predictions. Experiments suggest that 2Be3-Net has an improved ability to exploit spatial information compared to 2D or 3D only CNN solutions. We also evaluate our network on the prediction of clinical outcomes of head-and-neck cancer. The proposed pipeline outperforms PET radiomics approaches on the prediction of loco-regional recurrences and overall survival. Innovative deep learning architectures combining a pre-trained network with a 3D CNN could therefore be a great alternative to traditional CNN and radiomics approaches while empowering small and medium-sized datasets. 

Ronan Thomas, Elsa Schalck, Damien Fourure, Antoine Bonnefoy and Inaki Cervera-Marzal, 2Be3-Net : Combining 2D and 3D convolutional neural networks for 3D PET scans predictions, Proc. of the 2nd International Conference on Medical Imaging and Computer-Aided Diagnosis, 2021.

Watch Ronan present the paper on YouTube.

Click here to access the paper in its preprint form.

Releated Posts

Calibrate to Interpret

Trustworthy machine learning is driving a large number of the ML community works in order to improve ML acceptance and adoption. In this paper, we show a first link between uncertainty and explainability, by studying the relation between calibration and interpretation.
Read More

Mass Estimation of Planck Galaxy Clusters using Deep Learning

Galaxy cluster masses can be inferred indirectly using measurements from X-ray band, Sunyaev-Zeldovich (SZ) effect signal or optical observations. Unfortunately, all of them are affected by some bias. Alternatively, we provide an independent estimation of the cluster masses from the Planck PSZ2 catalogue of galaxy clusters using a machine-learning method.
Read More