A self-supervised learning based method for fish disease diagnosis utilizing extremely limited samples
Liu, Zehui Ye, Shanshan Li, Xuanzhi Luo, Xianwu Pan, Jiaji He, Miaolei Qin, Wei Xiao, Jun Cai, Yaoyi Shen, Tiantian
Liu, Zehui
Ye, Shanshan
Li, Xuanzhi
Luo, Xianwu
Pan, Jiaji
He, Miaolei
Qin, Wei
Xiao, Jun
Cai, Yaoyi
Shen, Tiantian
Supervisor
Department
Computational Biology
Embargo End Date
Type
Journal article
Date
License
Language
English
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
The extreme scarcity of high-quality annotated datasets in digital aquaculture severely hinders the application of deep learning-based fish disease diagnostics. In addition, general-purpose pretrained models may be less sensitive to the fine-grained pathological cues required for fish disease diagnosis. To address these challenges, this study proposes a fine-grained Masked Autoencoder (MAE) framework. This method aims to uncover the intrinsic structure within unlabeled domain data for robust fish disease diagnosis utilizing extremely limited samples. The framework first establishes an annotation-free preprocessing pipeline combining Grounding-DINO and the Segment Anything Model (SAM) to precisely eliminate complex background noise without human intervention. Then, an adaptive multi-view hybrid loss function covering pixel, frequency, perceptual, and gradient domains is designed, enabling the model to effectively capture high-frequency, fine-grained pathological features during reconstruction tasks. By integrating Low-Rank Adaptation (LoRA), the model efficiently transfers pretrained knowledge to downstream tasks by updating only approximately 3% of its parameters, effectively mitigating overfitting risks under data scarcity conditions. Extensive experiments demonstrate that this framework exhibits significant performance advantages across various data scales. Particularly in the highly challenging one-shot scenario, its ROC-AUC and PR-AUC reach 82.1% and 82.8%, respectively. The performance exceeds general visual foundation models and traditional supervised learning models by over 14%. Results indicate that for extracting fine-grained fish pathological features, small-scale domain-specific pretraining yields more discriminative representations than large-scale general transfer learning. This study reveals a very promising strategy other than foundation models to address data scarcity for digital fish disease surveillance.
Citation
Z. Liu, S. Ye, X. Li, X. Luo, J. Pan, M. He , et al., "A self-supervised learning based method for fish disease diagnosis utilizing extremely limited samples," Computers and Electronics in Agriculture, vol. 250, pp. 111952-111952, 2026, https://doi.org/10.1016/j.compag.2026.111952.
Source
Computers and Electronics in Agriculture
Conference
Keywords
46 Information and Computing Sciences, 4611 Machine Learning, 6 Clean Water and Sanitation
Subjects
Source
Publisher
Elsevier