Rapid label-free identification of pathogenic bacteria species from a minute quantity 1 exploiting three-dimensional quantitative phase imaging and artificial neural network

The healthcare industry is in dire need for rapid microbial identification techniques. Microbial infection is a major 27 healthcare issue with significant prevalence and mortality, which can be treated effectively during the early stages 28 using appropriate antibiotics. However, determining the appropriate antibiotics for the treatment of the early stages 29 of infection remains a challenge, mainly due to the lack of rapid microbial identification techniques. Conventional 30 culture-based identification and matrix-assisted laser desorption/ionization time-of-flight mass spectroscopy are 31 the gold standard methods, but the sample amplification process is extremely time-consuming. Here, we propose 32 an identification framework that can be used to measure minute quantities of microbes by incorporating artificial 33 neural networks with three-dimensional quantitative phase imaging. We aimed to accurately identify the species 34 of bacterial bloodstream infection pathogens based on a single colony-forming unit of the bacteria. The successful 35 distinction between a total of 19 species, with the accuracy of 99.9% when ten bacteria were measured, suggests 36 that our framework can serve as an effective advisory tool for clinicians during the initial antibiotic prescription. at single CFU level was verified in a comparative experiment. Our proposed framework was compared with two other approaches utilizing the 2D equivalent


Introduction 42
Infection by microorganisms is one of the major healthcare issues worldwide, causing a significant number 43 of casualties and a large amount of healthcare expense. Bacteria notably account for a large portion of life-44 threatening infections. During the year 2015, bacterial infections caused 4.4 million deaths, among a total of 45 8.8 million casualties by infections of any etiology (Hessling et al., 2017). In addition, the cost for treating 46 bacterial infections accounts for 8.7% of the national health spedning in US (Torio and Moore, 2016). 47 The ideal treatment for an infection is the administration of appropriate antibiotics during the early stage. 48 However, this is not easily implemented in the clinical settings, owing to the difficulty in rapid determination 49 of the pathogen. Early prescriptions of antibiotics are commonly carried out empirically without the complete 50 understanding of the etiology, and thus are often imperfect (García, 2009;Peterson et al., 2014) as antibiotics 51 vary in the efficacy for different pathogens (Hutchings et al., 2019). A systematic review underlines that 46.5% 52 of sepsis patients were given inappropriate empricial antibiotic treament and suffered 1.6-fold increased 53 mortality risk (Paul et al., 2010). Accordingly, a rapid method for identifying the pathogen is required. 54 Conventional phenotypic approaches are time-consuming and often nonspecific, despite being relatively 55 simple to perform (Bizzini and Greub, 2010). Culture tests, biochemical tests, and microscopic examination 56 of gram-stained specimens are well-known conventional methods for microbial identification. They require 57 hours or days of incubation for the metabolic activity or growth to take place. Molecular diagnostic methods 58 are not scalable because of their process-specific sensitivity and high cost, even though they provide detailed 59 information (Bizzini and Greub, 2010). Notably, 16S ribosomal RNA sequencing and real-time polymerase 60 chain reaction offer genetic evidence regarding the identity of the pathogen. While these methods can 61 precisely screen for a specific pathogen, the effectiveness of detection relies on the experimental setting, such 62 as the choice of the primer or probe. Along with the relatively high cost, this technical intricacy limits the 63 applicability of the molecular diagnostic methods. 64 In recent years, matrix-assisted laser desorption/ionization time-of-flight mass spectroscopy (MALDI-TOF 65 MS) has become the gold standard for microbial identification (Bizzini and Greub, 2010;Seng et al., 2009), 66 owing to its robust capability to investigate the molecular profile of the specimen. However, MALDI-TOF 67 MS typically entails a turnaround time above 24 h, since sample amplification must precede to secure a 68 detectable level of signal (Lin et al., 2018). A previous study indicated that a minimum of 10 5 colony-forming 69 units (CFUs) are required for MALDI-TOF MS-based detection of bacteria (Barreiro et al., 2017). In clinical 70 . CC-BY 4.0 International license available under a not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (which was this version posted June 23, 2021. ;https://doi.org/10.1101https://doi.org/10. /596486 doi: bioRxiv preprint et al., 2016. 129 An individual 3D RI tomogram referred to the distribution of RI in a 12.8 μm × 12.8 μm × 12.8 μm volumetric 130 space, sampled at a voxel resolution of 100 nm × 100 nm × 200 nm. Each 3D RI tomogram contained a single 131 bacterium or several bacteria that were adherent together after fission; we term this qauntity a single CFU 132 henceforth (1) in conformity to the definition of CFU (Hazan et al., 2012;Krieger, 2010) and (2) to connote 133 the sample quantity required for our framework. A manual inspection of each 3D RI tomogram ensured that 134 noisy measurements were ruled out before establishing the database. 135 136 2.3. Artificial neural network 137 The structure of ANN utilized in this framework mainly consists of 3D convolutional operations which can 138 effectively explore the 3D structure of 3D RI tomograms. More specifically, the dense connections between 139 the convolutional operations induce the ANN to revisit the feature maps of the shallower layers even at the 140 deep layers. Fig. 2 illustrates the structure of our ANN in detail. The structure is inspired by the convolutional 141 ANN design that outperformed most of the other designs in the benchmark tasks of 2D image analysis (Huang 142 et al., 2017). The four dense blocks include 12, 24, 64, and 64 colvolutional operations, respectively, from 143 shallow to deep. The number of feature channels after the initial convolution and the growth rate of the feature 144 channels are 64 and 32, respectively. 145 The ANN was optimized by minimization of the cross-entropy loss between the ground truth and the 146 prediction. For each species, 40 tomograms were randomly chosen as the blind test dataset and another 40 147 tomograms were randomly chosen as the validation dataset. The remaining tomograms composed the training 148 dataset, which were directly reflected in the loss minimization process. The loss that occured in the training 149 dataset was reduced using the stochastic gradient descent algorithm. The step size of the stochastic gradient 150 descent algorithm was scheduled according to the cosine annealing method (Loshchilov and Hutter, 2016) at 151 an initial step size of 0.001 and a period of 64 epochs. During training, data augmentation took place for each 152 tomogram, once every epoch, to prevent overfitting of the trained model. The augmentation included 153 random processes of horizontal crop, horizontal rotation, and Gaussian noise. During the blind test, each input 154 tomogram was horizontally cropped around the center to provide an identical dimension. These processes 155 resulted in an input tomogram of 9.6 μm × 9.6 μm × 12.8 μm to be fed into the ANN. A single training epoch 156 through the entire training dataset took approximately 10 min, when using eight graphics processing units of 157 . CC-BY 4.0 International license available under a not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (which was this version posted June 23, 2021. ; https://doi.org/10.1101/596486 doi: bioRxiv preprint GeForce GTX 1080ti and a central processing unit of Xeon E5-2600. The ANN was trained for 2,000 epochs 158 while saving models that yielded high training accuracies or validation accuracies. The ANN and the 159 optimization were implemented using PyTorch 1.0.0. 160 The algorithm for the blind test involved the predictions of multiple best-performing ANN models. Models 161 with the highest accuracies for the training and validation datasets were chosen and integrated, to exploit a 162 wider variety of features and prevent model-by-model variance. In search of the optimal strategy for chosing 163 and integrating multiple models, four relevant parameters were explored. These parameters included the 164 number of integrated models weighting between the accuracies for the training validation dataset, whether or 165 not to normalize the neural activation, and the formula for integrating the predictions by the chosen models. 166 Four options were considered as the formula for integrating the predictions: taking the average, taking the 167 exponential average, voting, and taking the maximum projection of the neural activation. The combination 168 of the parameters which yielded the highest validation accuracy established the algorithm for the blind test. 169 170

Results 171
The key function of our framework is to assess the species of the bacterial pathogen under a quantity of a 172 single CFU level. 3D QPI and ANN classification can provide preliminary results during the early stages of 173 infections, whereas the results of gold standard methods will be available dozens of hours later. Incorporation 174 of our framework into the gold standard routine is practicable since our framework operates with a minute 175 quantity of bacteria without destroying nor chemically modifying the bacteria. 176 177 3.1. Three-dimensional images of the bacteria 178 A database, which comprised 10,556 3D RI tomograms, was established with 19 different species of BSI 179 pathogens. The 19 species accounted for around 90% of all BSI-related cases, as indicated by the annual data 180 from a 1,000-bed tertiary care institute (Opota et al., 2015). The 3D RI tomograms of the 19 species showed 181 that 3D QPI effectively conveys the microscopic structure of bacteria ( The risk of error could be reduced through a broader interpretation of the neural activaton. To be precise, 206 narrowing down a few species that display high neural activation achieved lower rate of missing the correct 207 species, compared to the single-species prediction. This approach secures additional sensitivity at the cost of 208 specificity, which is a strategic trade-off. Approximately 94.3% of the blind test data included the correct 209 species, two of which had the highest values of neural activation. The probability further increased to 97.1% 210 when considering the top three values of neural activation ( Fig. 4(b)). This considerable reduction of error 211 was due to the robust feature-extracting ability of our ANN; the ANN recognizes the features related to the 212 correct species, even in the cases of misidentified tomograms. 213 214 3.3. Error in species identification 215 . CC-BY 4.0 International license available under a not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (which was this version posted June 23, 2021. ; https://doi.org/10.1101/596486 doi: bioRxiv preprint To characterize the errors in each species, the blind test results for all 19 species were investigated using the 216 confusion matrix (Fig. 5(a) The distribution of the species with the second and third highest values of neural activation visualizes the 223 similarity projected by the ANN (Fig. 5(b)). In this distribution, it is evident that the ANN reflects the 224 morphological similarities between different species belonging to close categories. For instance, two clusters 225 representing bacilli and other bacteria can be outlined, whereas the similarity between gram-positive bacilli 226 further stands out compared with the similarity among all bacilli. 227 Classification tasks referring to the gram-stainability and respiratory metabolism were also carried out, 228 resulting in accuracies of 94.6% and 94.2%, respectively (Fig. 5(c) and (d) The accuracy of species identification significantly increased when multiple 3D RI tomograms were reflected 234 in the prediction. The neural activation averaged over the inferences of multiple tomograms displayed a high 235 probability to indicate the correct species, even for cases where most individual tomograms were 236 misidentified ( Fig. 6(a)). The error rate dropped more sharply than a simple reciprocal function of the number 237 of tomograms; 94.9% and 98.4% accuracy was achieved, respectively using two and three 3D RI tomograms 238 ( Fig. 6(b)). This dramatic gain in accuracy was attributable to the robustness of the trained ANN in extracting 239 species-related features, which was described in Section 3.2. In addition, a quantitative analysis underpinned 240 that the correct predictions were made with higher contrasts in the neural activation compared to the 241 mispredictions (Appendix A.3). This analysis explained how the averaging process selectively promotes 242 correct predictions while flattening the erroneous signals in the neural activation. 243 244 . CC-BY 4.0 International license available under a not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (which was this version posted June 23, 2021. ; https://doi.org/10.1101/596486 doi: bioRxiv preprint

Discussion 245
We propose a framework for species identification of bacterial pathogens at a single CFU level using 3D QPI 246 and ANN. The exceptionally high accuracy under a limited sample quantity is attributable to the remarkable 247 single-cell profiling ability of 3D QPI and the feature-extracting ability of ANN. Results show that the ample 248 species-related features in a 3D RI tomogram are robustly extracted by the trained ANN, overcoming the 249 quantity requirement of the previous methods. 250 We believe that the proposed framework will efficiently refine the initial antibiotics prescribed in the clinical 251 settings. Our species identification accuracy based on a single CFU of bacteria is comparable to the 252 performance of MALDI-TOF MS in identifying the species of blood-born pathogens (Drancourt, 2010). The 253 risk of misidentification by our framework can be suppressed by taking multiple species into consideration. 254 The risk of missing the correct pathogen dramatically dropped when two or three most likely species were 255 selected. Our framework is also capable of being flexibly tuned for broader categories of bacteria such as 256 gram-stainability or aerobicity. Even though these categories are not as specific as species, they can play a 257 vital role in guiding antibiotic prescriptions. For instance, gram-positive pathogens can be effectively treated 258 with vancomycin, whereas the gram-stainability can be determined by a destructive staining of sufficient 259 sample. In addition, the accuracy of our framework can sharply increase through additional measurements of 260 3D RI tomograms. This signifies that our framework can be more accurate depending on the available sample 261 quantity, compared to the baseline of the single-CFU performance. Furthermore, this framework can be 262 incorporated along with the routine methods of microbial identification, including MALDI-TOF MS. The 263 high performance at a minute sample quantity and the noninvasive property allow our framework to be added 264 without exhausting the limited quantity of the sample. 265 Future studies on sample processing will propel our framework towards more immediate use. A condition 266 suitable for imaging bacteria has to be met to perform our single-CFU level identification. For blood samples, 267 this condition is achieved by performing lysis centrifugation after the initial blood culture (Kirn and Weinstein, 268 2013). However, application of our framework before completion of the blood culture is possible if a high-269 throughput procedure for enrichment of bacteria is introduced. A prominent and practical technique is the 270 selective collection of particles utilizing advanced fluidic systems (Balyan et al., 2020;Kuntaegowdanahalli 271 et al., 2009;Lee et al., 2019;Lei et al., 2012). Bacteria have also been prominent targets of collection using 272 fluidic systems (D'Amico et al., 2017;Jung et al., 2020). The adoption of these strategies will facilitate the 273 . CC-BY 4.0 International license available under a not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (which was this version posted June 23, 2021. ; https://doi.org/10.1101/596486 doi: bioRxiv preprint identification of the pathogen earlier than that suggested in our demonstration. 274 Moreover, validation using a larger diversity of pathogens will provide insights into the scope of application. 275 We expect the proposed framework to be applicable to pathogens causing different classes of infections, such 276 as urinary tract infections and lower respiratory infections, which are partially covered in this study. In 277 addition, it is yet to be assessed whether the framework is capable of distinguishing strains resistant to 278 antibiotics. The emergence of drug-resistant strains has compromised the established convention of antibiotic 279 prescription, and the need to screen out resistant strains has also been highlighted (Chamieh et al., 2020;280 Hutchings et al., 2019;Shariati et al., 2020). Investigatig the performance in identifying bacterial strains and 281 ensuring a higher accuracy to screening resistant strains will be crucial for improving the proposed framework. 282 283

Conclusion 284
To the best of our knowledge, our study demonstrates an unprecedented distinction of live and unmodified 285 bacteria at a single CFU level among a wide range of species. With a single measurement of a bacterium or a 286 CFU, we achieved the blind test accuracy of 82.5%, 94.6%, and 94.2% for species, gram-stainability, and 287 aerobicity, respectively. With ten individual measurements, we achieved, the blind test accuracy of 99.9%, 288 98.9%, and 99.9% for species, gram-stainability, and aerobicity, respectively. Our accuracy based on a single 289 measurement, which is comparable to the identification rate of MALDI-TOF MS, is facilitated by the precise 290 3D measurement of bacteria through 3D QPI and the statistical utilization of the measurement through ANN. 291 We believe that the proposed framework will substantially augment the early countermeasures against 292 bacterial infections; identifying the pathogen without the delay of sample amplification can provide a shortcut 293 for administrating the appropriate antibiotics. We note that, in principle, the application of our framework can 294 be brought forward to briefly after the sample collection if integrated with an advanced sample processing 295 techniques. 296 297 298

Appendix A 299
A.1. Suitability of 3D QPI for species identification of bacteria 300 The benefit of 3D QPI in identification of bacteria at single CFU level was verified in a comparative 301 experiment. Our proposed framework was compared with two other approaches utilizing the 2D equivalent 302 . CC-BY 4.0 International license available under a not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (which was this version posted June 23, 2021. ; https://doi.org/10.1101/596486 doi: bioRxiv preprint of our 3D ANN structure. One approach is trained to identify the species from 2D amplitude and phase delay 303 maps, while the other is trained to identify the species from the sinogram composed of 2D amplitude and 304 phase delay maps in multiple illumination angles. The approaches based on 2D and sinogram data achieved 305 67.6% and 68.0% blind-test accuracy respectively after training and model incorporation identical to that of 306 the 3D ANN (Fig. A.1). The significant difference in the accuracy suggests that 3D holographic microscope 307 offers features to ANN in a more ostensive manner. 308 309 A.2. Suitability of ANN for classification of 3D RI tomograms 310 The performance of ANN in recognizing the 3D RI tomograms of bacteria was compared to that of a 311 conventional machine learning approach. For the comparison we apply the strategy of threshold-derived 312 feature extraction and run k-nearest neighbors (k-NN) classifications based on the extracted features, that is, 313 an approach that had effectively classified 3D RI tomograms of lymphocytes (Yoon et al., 2017) (Fig. A.2(a)). 314 Scanning the values of k from 1 to 40, the lowest and the highest accuracies of the k-NN were 27.4% and 315 32.2% for k = 2 and k = 30 respectively (Fig. A.2(b)). As specified from the comparison, our implementation 316 of ANN was more capable of recognizing species-related characteristics, compared to the machine learning 317 with handcrafted features. 318 319 A.3. Contrast of neural activation 320 The dramatic rise of identification accuracy based on multiple 3D RI tomograms was accounted for by the 321 feature-extracting ability of the ANN. A tendency appearing in the neural activation displays how the 322 prediction significantly benefits from multiple tomograms. The contrast of neural activation, defined as the 323 highest activation value divided by the sum of all other positive activation values (Fig. A.3(a)), was 324 significantly higher in the correctly identified cases than the misidentified cases (Fig. A.3(b)). The difference 325 displays how taking the average of neural activation from multiple 3D RI tomograms elevates the correct 326 element of the neural activation. CC-BY 4.0 International license available under a not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (which was this version posted June 23, 2021. ; https://doi.org/10.1101/596486 doi: bioRxiv preprint Holograms including both the phase delay and the amplitude is measured while altering the illumination angle 514 using the DMD. (C) The three-dimensional RI tomogram is acquired by integrating the sinogram into the 515 scattering potential via optical diffraction tomography, followed by an iterative regularization. 516 517 . CC-BY 4.0 International license available under a not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (which was this version posted June 23, 2021. ; https://doi.org/10.1101/596486 doi: bioRxiv preprint The structure of the dense blocks allows features of shallower layers to be revisited in deeper layers. A dense 523 block consists of a pair of Convs followed by concatenation of the feature map before the two Convs. In each pair 524 of Convs, the first one has 1 × 1 × 1 kernels, and the second one has 2 × 2 × 2 kernels; meanwhile, the stride is 1 525 × 1 × 1 for both Convs. (c) The transition units shift the scale of the feature extracted by convolution. The Conv 526 in each transition unit has 1 × 1 × 1 kernels and 1 × 1 × 1 stride. 527 528 529 . CC-BY 4.0 International license available under a not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (which was this version posted June 23, 2021. ; https://doi.org/10.1101/596486 doi: bioRxiv preprint

Fig. 3. Three-dimensional refractive index tomograms of bacterial bloodstream infection pathogens. 531
Representative tomograms addressed in out study are rendenred in three dimension. Each tomogram represents 532 an individual species of bacterial pathogens. Scalebar = 2μm 533 534 535 . CC-BY 4.0 International license available under a not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (which was this version posted June 23, 2021. ; https://doi.org/10.1101/596486 doi: bioRxiv preprint CC-BY 4.0 International license available under a not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint (which was this version posted June 23, 2021. ; https://doi.org/10.1101/596486 doi: bioRxiv preprint