Research on Lightweight Detection of Small Lesions in Medical Imaging Based on Convolutional Neural Network Optimization

DOI: https://doi.org/10.62517/jmhs.202605102

Author(s)

Xuyan Wang

Affiliation(s)

Zhejiang Gongshang University, Hangzhou, Zhejiang, China

Abstract

In response to the core contradiction in the current medical imaging small lesion detection of convolutional neural network (CNN) models that "high precision leads to high consumption and lightweight leads to accuracy loss" [1], this paper proposes a "feature enhancement - lightweight collaborative optimization" strategy and a dynamic detection path to construct a lightweight small lesion detection model suitable for low-computation-power devices. This model is based on the Unet framework, strengthens the weak features of small lesions by embedding the Coordinate Attention attention mechanism, realizes model lightweighting by using Ghost convolution, and dynamically allocates computing power based on the lesion suspicion score - high-precision sub-networks are enabled for high-suspicion areas, and lightweight paths are adopted for low-suspicion areas. The experimental results on the LIDC-IDRI pulmonary nodule dataset and the ISLES stroke MRI dataset [2] show that the recall rate of small lesion detection in the proposed model reaches 86.3%, the accuracy rate reaches 91.2%, the number of parameters is controlled within 10M, and the reasoning time of a single image is ≤0.8 seconds. Compared with the traditional model of the same accuracy, the reasoning speed has been increased by 34.7%, effectively balancing the detection accuracy and efficiency. This research can provide technical support for real-time diagnosis of medical images in primary medical institutions and promote the downward allocation of medical resources.

Keywords

Convolutional Neural Network; Medical Imaging; Detection of Small Lesions; Lightweight; Precision and Efficiency Balance

References

[1] Guo Yansong, Cao Ying, Liu Wanyu, et al. Innovative Approach: A Cardiac Image-Assisted Diagnosis and Treatment System Based on Fused Neural Network [J] Internet of Things technology,2024,14(06):2. [2] Kshatri S S, Singh D. Convolutional neural network in medical image analysis: a review[J]. Archives of Computational Methods in Engineering, 2023, 30(4): 293-2810. [3] Lo SCB, Chan H P, Lin JS, et al. Artificial convolution neural network for medical image pattern recognition[J]. Neural networks, 1995, 8(7-8): 1201-1214. [4] Palwankar T, Kothari K. Real time object detection using ssd and mobilenet[J]. Int. J. Res. Appl. Sci. Eng. Technol, 2022, 10: 831-834. [5] Kshatri S S, Singh D. Convolutional neural network in medical image analysis: a review[J]. Archives of Computational Methods in Engineering, 2023, 30(4): 2793-2810. [6] Girshick R. Fast r-cnn[C]//Proceedings of the IEEE international conference on computer vision. 2015: 1440-1448. [7] Sanjay Kumar K K R, Subramani G, Thangavel S K, et al. A mobile-based framework for detecting objects using ssd-mobilenet in indoor environment[M]//Intelligence in Big Data Technologies—Beyond the Hype: Proceedings of ICBDCC 2019. Singapore: Springer Singapore, 2020: 65-76. [8] Lee S H. A Study on the Performance Evaluation of the Convolutional Neural Network–Transformer Hybrid Model for Positional Analysis[J]. Applied Sciences, 2023, 13(20): 11258. [9] Tan M, Pang R, Le Q V. Efficientdet: Scalable and efficient object detection[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2020: 10781-10790. [10] Wang Z, Li T. A lightweight CNN model based on GhostNet[J]. Computational intelligence and neuroscience, 2022, 2022(1): 8396550. [11] Zhang H, Liu M, Qi Y, et al. Efficient brain tumor segmentation with lightweight separable spatial convolutional network[J]. ACM Transactions on Multimedia Computing, Communications and Applications, 2024, 20(7): 1-19. [12] Zhang H, Liu M, Qi Y, et al. Efficient brain tumor segmentation with lightweight separable spatial convolutional network[J]. ACM Transactions on Multimedia Computing, Communications and Applications, 2024, 20(7): 1-19.