Deep Equilibrium Models (DEQs) have recently emerged as a promising approach to building implicit deep learning models that can achieve on-par accuracy with traditional explicit models but with considerably smaller sizes. However, the significant downside of DEQs is their slow inference speed, primarily due to the time cost of the fixed-point solver. This paper proposes to overcome this issue by applying HyperSolver, a novel technique that replaces traditional fixed-point solvers with a lightweight neural network. This is an extension of our previous work on PolypDeq concerning DEQs for medical image segmentation as an attempt to accelerate our existing implicit models. Experimental results show that our new models using Hyper-Solver can achieve similar results to existing DEQ models on several benchmark medical image datasets while having a significant speedup in inference time (about 9 times). To the best of our knowledge, this is the first attempt to accelerate DEQs for medical image segmentation using HyperSolver, representing a significant step towards making implicit deep learning models more practical for real-world applications.
Keyword
Semantic segmentation, polyp segmentation, implicit deep learning, deep equilibrium models
