[Paper Review] ASF-YOLO: A Novel YOLO Model with Attentional Scale Sequence Fusion for Cell Instance Segmentation

Min-Seong Kang, Chee‐Ming Ting+2 more|arXiv (Cornell University)|Dec 11, 2023

Cell Image Analysis TechniquesBiochemistry, Genetics and Molecular Biology45 references17 citations

TL;DR

ASF-YOLO introduces Attentional Scale Sequence Fusion to enhance multi-scale feature fusion for fast and accurate cell instance segmentation, achieving state-of-the-art metrics on a public dataset.

ABSTRACT

We propose a novel Attentional Scale Sequence Fusion based You Only Look Once (YOLO) framework (ASF-YOLO) which combines spatial and scale features for accurate and fast cell instance segmentation. Built on the YOLO segmentation framework, we employ the Scale Sequence Feature Fusion (SSFF) module to enhance the multi-scale information extraction capability of the network, and the Triple Feature Encoder (TFE) module to fuse feature maps of different scales to increase detailed information. We further introduce a Channel and Position Attention Mechanism (CPAM) to integrate both the SSFF and TPE modules, which focus on informative channels and spatial position-related small objects for improved detection and segmentation performance. Experimental validations on two cell datasets show remarkable segmentation accuracy and speed of the proposed ASF-YOLO model. It achieves a box mAP of 0.91, mask mAP of 0.887, and an inference speed of 47.3 FPS on the 2018 Data Science Bowl dataset, outperforming the state-of-the-art methods. The source code is available at https://github.com/mkang315/ASF-YOLO.

Motivation & Objective

• Advance cell instance segmentation by integrating multi-scale feature fusion with attention mechanisms.
• Develop modules to fuse scale sequences and enrich detail preservation in segmentation.
• Enhance detection and segmentation with channel and position attention to focus on informative features.

Proposed method

• Introduce Scale Sequence Feature Fusion (SSFF) to augment multi-scale information extraction.
• Incorporate a Triple Feature Encoder (TFE) to fuse feature maps across scales for richer details.
• Apply a Channel and Position Attention Mechanism (CPAM) to steer SSFF and TFE toward informative channels and spatial regions.
• Build on the YOLO segmentation framework to enable fast, end-to-end training and inference.
• Report performance on two cell datasets with metrics including box mAP and mask mAP.

Experimental results

Research questions

• RQ1 Can attentional scale sequence fusion improve multi-scale feature representation for cell instance segmentation?
• RQ2 Does combining SSFF, TFE, and CPAM yield superior segmentation accuracy and speed compared with existing YOLO-based methods?
• RQ3 What are the performance gains on standard cell segmentation benchmarks?

Key findings

• Box mAP of 0.91 on the 2018 Data Science Bowl dataset.
• Mask mAP of 0.887 on the same dataset.
• Inference speed of 47.3 FPS on the 2018 Data Science Bowl dataset.
• ASF-YOLO outperforms state-of-the-art methods on the tested cell datasets.
• The proposed modules (SSFF, TFE, CPAM) synergistically improve both detection and segmentation quality.