Human Brain-Inspired Network Using Transformer and Feedback Processing for Cell Image Segmentation

Abstract

Semantic segmentation of microscopy cell images by deep learning plays a crucial role in advancing medicine and cell biology research. We considered that Transformers, which have recently outperformed CNNs in image recognition, could also be improved and developed for cell image segmentation. Transformers tend to focus more on contextual information than on detailed information. This often results in blurred boundaries and difficulty in distinguishing small cellular structures. Hybrid models combining Transformers and CNNs have been proposed to address this issue, but they introduce high computational costs and architectural complexity. Therefore, to supplement or reinforce the missing information, we hypothesized that feedback processing in the visual cortex of the human brain could be highly effective. Our proposed Feedback Former is a novel architecture for semantic segmentation, inspired by the structure of the human brain. In the visual cortex of the human brain, the inference is made by feedforward processing from lower to upper layers, followed by the transfer of information in the reverse direction by feedback processing. In this process, specific information is emphasized or suppressed, and recognition is modified. This modification of recognition might occur in neural networks as well, so we incorporate feedback processing into a segmentation model that uses Transformers as an encoder. Feedback processing is implemented by directly connecting the output neighborhood of the model to the lower layers. Feeding back feature maps with detailed information near the output of the model obtained once inference is performed to the lower layers improves the accuracy of segmentation, especially in areas with complex textures and small objects, by enhancing the ability to extract features near boundaries and details. We further propose Lite Feedback Module, a computationally efficient alternative to conventional feedback modules. Unlike hybrid models that require additional CNN components, this module improves segmentation accuracy while maintaining lower computational costs. Experiments on three different cell image datasets confirmed that the proposed method surpasses the methods without feedback processing, demonstrating its superior accuracy in cell image segmentation. Our method achieved higher segmentation accuracy while consuming less computational cost than conventional feedback approaches. Our method enhanced accuracy more efficiently than simply increasing the model size or using a hybrid structure with CNNs.