Improving generalization in classification of novel bacterial strains: a multi-headed ResNet approach for microscopic image classification
Year: 2024
Type: article
Abstract: The purpose of this work is to design a system for microscopic bacterial images classification that can be generalized to new data. In the course of work, a dataset containing 23 bacterial species was collected. We use a strain-wise method for dividing the dataset into training and test sets. Such splitting (in contrast to random division) allows evaluating the performance of classifiers on new strains in the case of intra-species visual variability of bacteria. We propose a "Multi-headed" ResNet (ResNet-MH) for the analysis of microscopic images of bacterial colonies. This approach forces the neural network to analyze features of different resolutions, such as the shape of individual bacterial cells and the shape and number of bacterial clusters during training. Our network achieves the 41.6% accuracy species-wise and 64.06% accuracy genera-wise. The proposed method of dataset splitting guarantees generalization to new unseen strains, whereas random splitting into training and test sets leads to overfitting of the system (accuracy is over 90%). For the 10 visually strain-wise stable species, the accuracy of the proposed system reaches 83.6% species-wise.
Source: CyberLeninK (CyberLeninka)
Authors: Valeria Olegovna Yachnaya, Maria Anatolievna Mikhalkova, Roman Olegovich Malashin, Vadim Rostislavovich Lutsiv, Lyudmila Aleksandrovna Kraeva, Galina Nikolaevna Khamdulayeva, Vitaliy Evgenievich Nazarov, Vladimir Petrovich Chelibanov
Language: English
Subfield: Biophysics
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W6888005355
FWCI:
Cites:
Cited by:
Related to: