Deep Learning based Disease Prediction Using External and Internal Images of Paralichthys Olivaceus

Early detection and rapid response to fish disease are essential because fish farms are vulnerable to disease. However, since conventional fish disease diagnosis relies on expert knowledge, it takes a lot of time and money, raising the mortality risk. Recently, to prevent this, research on fish disease prediction using deep learning has been actively conducted. This study proposes a method to diagnose fish diseases by predicting disease symptoms using a deep-learning classification model and an object-detection model on internal and external images of Paralichthys olivaceus. First, images of internal organs are extracted from internal images of Paralichthys olivaceus using a deep-learning segmentation model, and internal symptoms are predicted by training the deep-learning classification model. Next, a deep-learning object-detection model is trained on external images of the Paralichthys olivaceus to detect external symptoms. The internal and external symptoms are then merged and used to predict diseases. We compared the performance of symptom weight-based, data statistics-based, and symptom set learning-based methods in disease prediction. The proposed internal-symptom classification by organ was superior to the symptom object detection, and predicting disease by learning the symptom set using deep learning achieved the highest performance with Top-1 accuracy of 0.5818, and Top-3 accuracy of 0.8364.

鱼类疾病的早期检测与快速响应至关重要，原因在于水产养殖场极易受到病害侵袭。然而传统鱼类疾病诊断依赖专业人员的专业知识，不仅耗时耗财，还会提升养殖鱼类的死亡率风险。近年来，为规避此类问题，基于深度学习的鱼类疾病预测研究已得到广泛开展。本研究针对褐牙鲆（Paralichthys olivaceus）的体内与体外图像，提出了一种结合深度学习分类模型与目标检测模型、通过预测疾病症状来实现鱼类疾病诊断的方法。首先，借助深度学习分割模型从褐牙鲆的体内图像中提取内脏器官图像，并通过训练深度学习分类模型预测其体内症状；随后，针对褐牙鲆的体外图像训练深度学习目标检测模型，以识别其体外症状。随后将体内与体外症状进行融合，用于疾病预测。本研究对比了基于症状权重、基于数据统计以及基于症状集学习的三类疾病预测方法的性能。结果表明，本研究提出的按器官分类的体内症状识别方法，性能优于症状目标检测方法；而通过深度学习学习症状集进行疾病预测的方案性能最优，其Top-1准确率达0.5818，Top-3准确率达0.8364。