PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Ming Kang, Chee-Ming Ting, Fung Fung Ting, Raphaël C. -W. Phan

arXiv:2312.06458v2cs.CVeess.SPstat.AP

TL;DR

Real-time cell instance segmentation remains challenging for dense, small cells. ASF-YOLO fuses multiscale features and attention within YOLO, achieving best reported performance on the evaluated datasets, including 0.91 Box mAP50, 0.887 Mask mAP50, and 47.3 FPS.

  • Problem

    Existing CNN-based methods provide sub-optimal real-time cell instance segmentation for dense, small cells, while feature pyramids inadequately exploit correlations across scales.

  • Method

    ASF-YOLO extends YOLO with SSFF and TFE multiscale fusion modules plus channel and position attention to improve small-cell segmentation.

  • Results

    0.91 Box mAP50, 0.887 Mask mAP50, and 47.3 FPS were achieved on the evaluated dataset, with best performance reported across the cell datasets.

  • Takeaways & Limitations

    ASF-YOLO provides accurate and fast cell instance segmentation across cell images with varying cell types, balancing detection accuracy and computational speed.

  • Takeaways & Limitations

    Further accuracy improvement is needed while maintaining segmentation efficiency, and the CPAM mechanism may slightly increase computational effort.

Abstract

from arXiv · show

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.

1. Introduction

The introduction motivates cell instance segmentation as a difficult, detail-sensitive problem involving small, dense, overlapping cells and blurred boundaries. It presents ASF-YOLO as a one-stage YOLOv5-based model that improves small-object segmentation through multiscale feature fusion and attention mechanisms.

  • Cell image analysis supports important applications in medicine and cell biology, while CNNs can learn characteristic information across different cell images.
  • YOLO-based instance segmentation offers strong accuracy and speed, but cell images remain challenging because objects are small, dense, overlapping, and bounded ambiguously.
  • ASF-YOLO extends YOLOv5 for cell instance segmentation by adding multiscale feature fusion and attention mechanisms, especially in the neck.
  • The framework uses a CSPDarknet53 backbone to extract multidimensional feature information from cell images before neck-level fusion and attention processing.
  • SSFF and TFE fuse backbone feature maps within a PANet structure, with SSFF normalizing, upsampling, and concatenating multiscale features before 3D convolution.

2. Related work

Related work has advanced cell instance segmentation through deep-learning-based detection and segmentation methods, including Mask R-CNN and single-shot detectors. Recent YOLO-based efforts emphasize attention mechanisms, feature extraction, architectural changes, loss functions, contour regression, and hybrid semantic-segmentation designs.

  • Cell instance segmentation: Deep learning has improved automated nucleus segmentation, while cell instance segmentation additionally supports completing cell counting beyond semantic segmentation.Several studies proposed simultaneous cell detection and segmentation based on Mask R-CNN.
  • YOLO-based methods: Recent YOLO instance-segmentation improvements focus on attention mechanisms, improved backbones or networks, and loss functions.Examples include integrating SENet into YOLACT and adding CBAM to YOLO-based models.
  • YOLO-based methods: Hybrid YOLOMask and YUSEG models combine optimized YOLO architectures with U-Net semantic segmentation to improve instance-segmentation accuracy.Another approach enhances masks through explicit contour regression using polar distance and sector losses.

3. The proposed ASF-YOLO model

ASF-YOLO combines SSFF for multiscale semantic information, TFE for local fine details, and CPAM for channel and spatial attention to improve small-cell instance segmentation. Its design also replaces CIoU with EIoU in the detection head to improve regression.

  • Feature fusion architecture: ASF-YOLO combines SSFF multiscale features with TFE local details for small-object cell segmentation.SSFF captures high-level information across scales, while TFE captures local fine details.
  • CPAM module: CPAM integrates TFE and SSFF features through channel attention and position attention to emphasize informative channels and refine cell localization.The channel network uses global average pooling, fully connected layers, and a Sigmoid function; the position network combines channel-attended and SSFF outputs.
  • SSFF module: SSFF builds scale-sequence representations from P3, P4, and P5 using Gaussian smoothing, resolution alignment, depth concatenation, and 3D convolutions.P4 and P5 are channel-adjusted to 256 and resized to P3 resolution before scale-sequence extraction.
  • TFE module: TFE addresses densely overlapping small objects by encoding large-, medium-, and small-scale feature maps before concatenating them.Its output matches the medium-scale resolution and has three times the medium-scale channel number.
  • Detection head: EIoU replaces CIoU in the detection head to accelerate prediction-box convergence and improve regression accuracy.EIoU incorporates Euclidean center-point distance and width- and height-related terms.

4. Experiments · 4.1. Datasets

The experiments evaluate ASF-YOLO on the DSB2018 and BCC cell image datasets, covering varied imaging conditions and breast cancer histopathology. DSB2018 contains 670 annotated nuclei images, while BCC contains 160 images divided into training and test sets.

  • 4. Experiments: ASF-YOLO was evaluated on the DSB2018 and BCC cell image datasets.These experiments assess the proposed model across two cell-image benchmarks.
  • 4.1. Datasets: DSB2018 contains 670 cell nuclei images with segmented masks.The dataset is used to assess generalizability across cell type, magnification, and imaging modality variations.
  • 4.1. Datasets: Each DSB2018 mask contains one nucleus, with no overlap between masks.No pixel belongs to two masks in this dataset.
  • 4.1. Datasets: The BCC dataset contains 160 hematoxylin-and-eosin-stained histopathology images for breast cancer cell detection.The images include associated ground-truth data and were collected from UCSB CBI.
  • 4.1. Datasets: Performance comparisons for cell instance segmentation are reported separately for the DSB2018 and BCC datasets.Table 1 covers DSB2018, and Table 2 covers BCC; the best results are indicated in bold.
  • 4.1. Datasets: 128 BCC images form the training set, while 32 images form the test set.The partition assigns 80% of the images to training and 20% to testing.

4.2. Implementation details

Experiments used a fixed hardware and software setup with pretrained COCO initialization, 640 × 640 inputs, batch size 16, and 100 training epochs. The model was optimized with SGD using momentum 0.9 and an initial learning rate of 0.001.

  • 4.2. Implementation details: Experiments ran on an NVIDIA GeForce 3090 (24G) GPU with Pytorch 1.10, Python 3.7, and CUDA 11.3.The model used initial weights pretrained on the COCO dataset.
  • 4.2. Implementation details: 0.9 momentum and 0.001 initial learning rate were used with SGD optimization.The passage describes these hyperparameters as those of SDG.

4.3. Quantitative results

ASF-YOLO delivered the strongest reported DSB2018 performance, combining high segmentation accuracy with 47.3 FPS, and also achieved the best instance segmentation performance on the BCC dataset. These results indicate generalization across datasets with varying cell types.

  • DSB2018 results: 0.91 Box mAP50 and 0.887 Mask mAP50 were achieved with 46.18 million parameters, alongside an inference speed of 47.3 FPS on DSB2018.These were reported as the best performance in the comparison.
  • DSB2018 results: ASF-YOLO surpassed the classical one-stage algorithms SOLO and YOLACT on the DSB2018 dataset.The comparison included Mask R-CNN, Cascade Mask R-CNN, SOLO, SOLOv2, YOLACT, Mask RCNN Swin T, YOLOv5l-seg v7.0, and YOLOv8l-seg.
  • BCC results: ASF-YOLO achieved the best instance segmentation performance on the BCC dataset, demonstrating generalization to different datasets with varying cell types.The BCC result was presented as experimental validation of the model’s generalization ability.

4.4. Qualitative results

Qualitative comparisons on DSB2018 show that ASF-YOLO handles both dense, small-cell images and large-cell images with complex backgrounds. The TFE and SSFF modules contribute to recall and segmentation accuracy in these respective settings.

  • 4.4. Qualitative results: Fig. 6 visually compares ASF-YOLO with different methods on sample DSB2018 cell images.The comparison covers qualitative cell-segmentation results.
  • 4.4. Qualitative results: ASF-YOLO achieved good recall for images containing dense and small objects in a single channel.This result is attributed to the TFE module’s improved small-object detection performance.
  • 4.4. Qualitative results: ASF-YOLO provided good segmentation accuracy for large-sized cell images under complex backgrounds.This result is attributed to the SSFF module’s enhanced multiscale feature extraction performance.

4.5. Ablation study

The ablation studies evaluate ASF-YOLO’s proposed modules, attention mechanism, and backbone choice. Soft-NMS, EIoU loss, SSFF, TFE, and CPAM improve segmentation performance, while CPAM outperforms alternatives and the YOLOv5 backbone remains preferable to a C2f-based replacement.

  • Module contributions: Soft-NMS addresses mutual-occlusion errors among dense small cells, while EIoU loss improves small-object bounding boxes and increases mAP50:95 by 1.8%.The ablation attributes additional performance improvements to the SSFF, TFE, and CPAM modules for small-object instance segmentation.
  • Attention mechanisms: CPAM provides better performance than SENet, CBAM, and Coordinate Attention despite a slight increase in computation and parameters.Table 4 evaluates the compared attention mechanisms.
  • Attention mechanisms: CPAM extracts richer channel and positional feature information from original images than the other evaluated attention modules.Qualitative visualizations on the DSB2018 dataset compare segmentation results produced by different attentional modules.
  • Backbone selection: Replacing YOLOv5 C3 modules with YOLOv8 C2f modules in the backbone decreases performance on both datasets.The comparison supports selecting the YOLOv5 backbone for the improved neck configuration.

5. Conclusion

ASF-YOLO is presented as an accurate, fast cell instance-segmentation model that fuses spatial and scale features through SSFF, TFE, and CPAM. The authors note a balance between accuracy and speed, while identifying opportunities to improve accuracy and preserve segmentation efficiency.

  • Contributions: ASF-YOLO fuses spatial and scale features for cell-image detection and segmentation within the YOLO framework.SSFF and TFE enhance multiscale and small-object instance segmentation, while CPAM mines their feature information.
  • Performance and limitations: ASF-YOLO balances detection accuracy and computational speed, although CPAM may slightly increase computational effort.The authors emphasize maintaining segmentation efficiency for practical clinical implementation.
  • Future work: Future work could use dilated convolution, backbone transfer learning, and Transformers to improve contextual features, histological-feature extraction, and attention mechanisms.Dilated convolution is proposed without increasing computational effort or parameter count.

CRediT authorship contribution statement

The authors contributed across writing, technical development, validation, supervision, and project support. Ming Kang led the technical and analytical work, while the other authors contributed primarily through writing, validation, and oversight.

  • Author contributions: Ming Kang contributed to writing, visualization, validation, software, investigation, formal analysis, and conceptualization.He also handled both original-draft writing and review and editing.
  • Author contributions: Chee-Ming Ting contributed to writing, validation, supervision, project administration, funding acquisition, and conceptualization.His writing contributions included both review and editing.
  • Author contributions: Fung Fung Ting contributed to writing, validation, and supervision.The listed writing contribution was review and editing.
  • Author contributions: Rapha¨el C.-W. Phan contributed to writing, validation, and supervision.The listed writing contribution was review and editing.

Data availability

The authors state that a link to their data is provided in the References section.

  • Data availability: A data link is shared in the References section.
Loading 2312.06458v2…