Prediction of Multiple Microbes Present in Single Image Using Advanced Deep Learning Models
Ball Mukund Mani Tripathi, Nagothu Hemanth Nirmal Kumar, Shaik Khasim Sharif, Kudenti Sai Chand, Routu Santosh · 2024
The need for efficient and accurate methods of identifying specific microbes within complex communities has grown due to the vast amount of microbiome data available. This study focuses on multiple microbe identification within a given sample image using convolutional neural networks (CNNs). The CNNs have shown incredible capacity to extract insights from microbiome data, resulting in unique patterns and characteristic recognition for targeted microbes and accurate classification. The analysis for this study employs three CNN models, Googlenet, Resnet-50, and Efficientnet-b0, trained with different algorithms. The Efficientnet-b0 model trained with the SGDM algorithm performs the best with a smaller number of iterations(140) and better training accuracy(100%) and loss(0.0897) with testing accuracy(64.9123 %). But overall, the testing accuracy should be considered as a key metric according to that the Resnet-50 & Googlenet trained with ADAM got the highest accuracy (75.4386%) than others but among them, the Resnet-50 should be considered as the best model due to its training loss (0.0372) is lower than Googlenet and got average testing Precision, Recall & Fl Score as 0.6749, 0.7141 & 0.6617 respectively. The study emphasizes the importance of data preprocessing, model architecture selection, and training strategies tailored to the microbial identification task. Additionally, the versatility of CNN models across different microbiome datasets enhances their transferability.