# FC-YOLO: a fast inference backbone and lightweight attention mechanism-enhanced YOLO for detecting gastric adenocarcinoma in pathological image

## TL;DR

FC-YOLO is a fast and efficient AI model designed to detect gastric adenocarcinoma in histopathological images, improving diagnostic accuracy and speed.

## Contribution

FC-YOLO introduces a lightweight attention mechanism and efficient backbone for fast and accurate gastric adenocarcinoma detection.

## Key findings

- FC-YOLO achieved 82.8% mAP on a public dataset, outperforming YOLOv11s by 2.6%.
- The model reached 131.56 FPS inference speed while maintaining high accuracy.
- FC-YOLO demonstrated strong generalization with 85.7% mAP on an independent clinical dataset.

## Abstract

Gastric adenocarcinoma (GAC) is a leading cause of cancer-related mortality, but its histopathological diagnosis is challenged by image complexity and a shortage of pathologists. While deep learning models show promise, many are computationally demanding and lack the fine-grained feature extraction necessary for effective GAC detection.

We propose FC-YOLO, an optimized object detection framework for GAC histopathological image analysis. Based on the YOLOv11s architecture, FC-YOLO incorporates a FasterNet backbone for efficient multi-scale feature extraction, a lightweight Mixed Local-Channel Attention (MLCA) mechanism for feature recalibration, and Content-Aware ReAssembly of FEatures (CARAFE) for enhanced upsampling. The model was evaluated on a public dataset comprising 1,855 images and on a separate, independent clinical dataset consisting of 2,500 pathological images of gastric adenocarcinoma.

On the public dataset, FC-YOLO achieved a mean Average Precision (mAP) of 82.8%, outperforming the baseline YOLOv11s by 2.6%, while maintaining a high inference speed of 131.56 FPS. On the independent clinical dataset, the model achieved an mAP of 85.7%, demonstrating strong generalization capabilities.

The lightweight and efficient design of FC-YOLO enables superior performance at a low computational cost. It represents a promising tool to assist pathologists by enhancing diagnostic accuracy and efficiency, particularly in resource-limited settings.