Incomplete wing classifier

We curated a data set consisting of a class of complete wing images containing all landmarks and a second class consisting of incomplete wing images missing landmarks 4 or 6. Since manually finding training samples is time consuming, we aided the process by using the biological data set (variables lmkr and lmkl) and then used visual inspection to filter a set of training images. A data-centric approach was employed, focusing on consistency in labelling, such that the types of wings we introduced into the training set were not ambiguous and fell into well-defined groups. We removed examples that did not clearly fall into either class. The resulting data set consists of an even class distribution of incomplete and complete wing images. Images are labelled 1 for incomplete wings and 0 for complete wings. In total, we labelled 1227 images.

We compared three modern computer vision models with transfer learning that were used similarly for classifying insect wings in a previous study [35]. These models are ResNet18 [52], Inception [53], and VGG16 [54] with batch normalisation. We replaced the final fully connected layers, after the convolutional layers, with a fully connected layer of size 1. We then applied a sigmoid activation function, with an output in the range (0, 1).

A 3-channel input was used for all models. Input image size of 244 × 244 was used for both VGG16 and ResNet18. For the Inception model, we used an input image size of 299 × 299 since the default kernel size of 3 × 3 and a stride of 2 with zero paddings require the image dimensions to be odd numbers. We fine-tuned each classifier with unfrozen weights for 30 epochs with a batch size of 50. The model was saved at the lowest validation score in each training session. A learning rate of 0.0001 was used with the Adam optimiser and a binary cross-entropy loss function.

Evaluation: For each classifier, we obtained 95% confidence intervals for specificity, sensitivity, precision, f1 score, and accuracy by bootstrapping the predictions on the test set, randomly sampling 205 samples, 5000 times to create sample statistic distributions.