Single and multi-analyte deep learning-based analysis framework for class prediction in biological images

Abstract

Measurement of biological analytes, characterizing flavor in fruits, is a cumbersome, expensive and time-consuming process. Fruits with higher concentration of analytes have greater commercial or nutritional values. Here, we tested a deep learning-based framework with fruit images to predict the class (sweet or sour and high or low) of analytes using images from two types of trees in a single and multi-analyte mode. We used fruit images from kinnow (n = 3,451), an edible hybrid mandarin and neem (n = 1,045), a tree with agrochemical and pharmaceutical properties. We measured sweetness in kinnows and five secondary metabolites in neem fruits (azadirachtin or A, deacetyl-salannin or D, salannin or S, nimbin or N and nimbolide or E) using a refractometer and high-performance liquid chromatography, respectively. We trained the models for 300 epochs, before and after hyper-parameters evolution, using 300 generations with 50 epochs/generation, estimated the best models and evaluated their performance on 10% of independent images. The validation F1score and test accuracies were 0.79 and 0.77, and 82.55% and 60.8%, respectively for kinnow and neem A analyte. A multi-analyte model enhanced the neem A models prediction to high class when the D:N:Ss combined class predictions were high:low:high and to low class when D:Ns combined class predictions were low:high respectively. The test accuracy increased further to ~70% with a 10-fold cross-validation error of 0.257 across ten randomly split train:validation:test sets proving the potential of a multi-analyte model to enhance the prediction accuracy, especially when the numbers of images are limiting.