Abstract

Sithin Thulasi Seetha^1,2, Kurt Driessens³, Henry Woodruff¹, Tiziana Rancati⁴, Elena Bertocchi⁴, Ugo Pastorino², Philippe Lambin¹

¹GROW - School for Oncology and Developmental Biology, Maastricht University Medical Centre, Department of Precision Medicine, Maastricht, The Netherlands; ²IRCCS Foundation National Cancer Institute, Department of Thoracic Oncology, Milan, Italy; ³Maastricht University, Data Science & Knowledge Engineering, Maastricht, The Netherlands; ⁴IRCCS Foundation National Cancer Institute, Prostate Cancer Unit, Milan, Italy

To propose a novel fast, compact, and low compute hybrid neural architectural search (NAS) algorithm which we call gNEAT for complex deep learning applications; validate the algorithm, and apply it for the task of lung cancer segmentation.

gNEAT essentially searches over a space of possible network solutions starting from the most basic architecture (containing only the input and output nodes). During the course of the search, additional hidden nodes and connections are introduced incrementally. Compactness in the context of ANNs is measured based on the minimum number of parameters (nodes and connections) needed to solve a task, in relation to some known optimal solution. Validation of gNEAT's compactness is done by XOR and Autoencoder (AE) tasks (see Fig 1.a). Another important quality of gNEAT is that it can perform complex NAS with incredible speed and low computational cost. This is achieved by reducing the population size while ensuring diversity within the population for sufficient search space exploration. Furthermore, to offset the complexity associated with gradient-descent training, computation saving techniques such as the use of a proxy dataset (using resized dataset) and a blueprint (using a predefined structure and searching its components) are integrated (see Fig 1.c). Once the components are evolved, the resultant model is then applied to the original dataset. gNEAT’s abilities for complex NAS are showcased by evolving a 2D UNet architecture for lung cancer segmentation (see Fig 1.b) on the Medical Segmentation Decathlon (Lung D6) dataset. The results are then compared to the state-of-the-art (STOA) nnUNet pipeline (re-implemented). Even though nnUNet 3D holds the current STOA benchmark on this dataset, we will be limiting all the experiments to nnUNet 2D architecture to reduce complexity. In addition to this, we perform an external validation of nnUNet 2D and the evolved solution on the open-source NSCLC Radiomics (Lung 1) dataset.

The table below and figure 2 illustrates validation and application results. Moreover, gNEAT was able to return a solution within 5.66 ± 2.47 hours, averaged across 5 runs, on a single NVIDIA GTX 1080 Ti GPU using less than 4GB of GPU memory. 

gNEAT can replace the expensive, time-consuming, and exhaustive trial-and-error process of manual architectural search. Unlike similar automated NAS algorithms, gNEAT can return domain-specific compact solutions in a matter of hours on a low compute system accessible to the majority of research institutions.