Abstract
Reliable ovarian maturity assessment can support harvest decisions in cultured female mud crabs (Scylla paramamosain), but evaluating internal ovarian development non-destructively remains challenging. This study developed and evaluated a convolutional neural network (CNN)-based framework for binary ovarian maturity classification from smartphone-acquired RGB images without handcrafted feature extraction. Classification performance was systematically compared across two imaging modes: dorsal transillumination imaging (DTI) and flashlight-assisted imaging (FAI). The dataset comprised 2000 images (1000 per mode), with reference labels assigned by expert consensus. Five CNN architectures were evaluated under matched-condition, cross-condition, and combined-development settings, with five runs per architecture and development setting. Under matched conditions, VGG16 achieved the highest mean accuracy for DTI (0.9933), whereas DenseNet121 achieved the highest for FAI (0.9893). Cross-condition accuracies ranged from 0.7187 to 0.8947 for DTI→FAI and from 0.6907 to 0.8067 for FAI→DTI. Under combined DTI+FAI development, the highest mean accuracies were 0.9920 on DTI (VGG16) and 0.9947 on FAI (MobileNetV2). Combined development yielded higher mean accuracy than cross-condition evaluation on the same test subset for all five architectures. These findings characterize CNN performance across the evaluated imaging modes and may inform future non-destructive ovarian maturity assessment tools.
Funding Information
  • National Science, Research and Innovation Fund (ENG67012405)

This publication has 38 references indexed in Scilit: