While most works reported so far on the topic of tissue imaging with TPEF and SHG have dealt with images collected at optical resolution limited by diffraction, in GASTODEEP, besides conventional arhitectures for Non-linear Optical Imaging (NLO), we also tackle the topic of super-resolution NLO data at sub-diffraction resolution, based on the Re-scan concept, initially introduced for fluorescence based confocal imaging, and later on extended for NLO imaging in the H2020 aTTRACT: HARMOPLUS. Taking advantage that Re-scan Second Harmonic Generation Microscopy (Re-scan SHG) provides similar advantages in terms of resolution and sensitivity as RCM, we endeavour in GASTRODEEP to to perform Re-scan SHG imaging of fixed human tissue samples, prepared for conventional histopathology exam. Our main aim is to asess whether the increased resolution avaialble in the Re-scan configuration provides additional advantage in the automated assessment of tissues states by Deep Learning. A side result of our efforts on Re-scan SHG imaging of tissues, relates to the fact that RCM units adapted for Re-scan SHG can be easily coupled to optical microscopes available in clinical settings, and thus histopathology labs can immensely benefit from this development, as it can enable modern histopathology workflows complementing their traditional assays. In GastroDeep, we focus on showing that by resolving the collagen architecture at super-resolution histopathologists can identify subtle modifications that precede or develop during cancers, which can be of great help in refining their diagnostic, as confirmed by our research partners at the Carol Davila University of Medicine and Pharmacy. 

Mosaic. diffraction limited vs. Re-scan SHG images collected on fixed human tissue. Scale bar: 5μm.

Also related to the Re-scan concept, the GASTRODEEP project has performed work on fluorescence imaging of bacterial pathogens that have implications with respect to gastric cancer, which is the pathology at the focus of the project. Our ultimate goals are to correlate how bacteria in patients with gastric cancers respond to different types of antibiotics, and also to identify correlations between the onset and progressions of the diseases and infection by various species of pathogenic bacteria. While in the future we plan to address Helicobacter Pylori, the bacterial species that is known to have most implications with respect to gastrointestinal cancers, in the first steps we have addresss Acinetobacter baumannii, a nosocomial pathogen that represents an increasing global health threat, especially for the emergence of multidrug resistant (MDR) strains, which is also known to be correlated with gastric cancers. For this bacterial species, and most others, an accurate analysis of morphological features is important for assessing the mode of action and effectiveness of antimicrobial drugs, as those affecting cell wall synthesis or damaging the cell membrane. To this end we analysed the effects of colistin (polymyxin E), which now represents a last line, life-saving therapeutic option for treating infections caused by MDR and pan-resistant A. baumannii. Colistin is a cationic polypeptide whose antimicrobial activity is exerted by targeting the bacterial membrane. Colistin interacts with the outer membrane, a distinct feature of Gram-negative bacteria, leading to the displacement of divalent cations that stabilize the outer membrane, thus causing derangement of the cell membrane, leakage of intracellular contents and ultimately cell death. Our experiments showed that RCM can successfully be used to accurately characterize the membrane of A. baumannii cells, which were stained with SYTO9 and KK114, as discussed next. The reason behind our interest in analysing the filamentous cells of A. baumannii by RCM, relates to the fact that filamentation is supposed to be one of the mechanisms contributing to antibiotic resistance: when treated with sub-inhibitory concentration of certain classes of antimicrobials, bacterial cells undergo asymmetric divisions at the filament tips, giving rise to budded resistant offspring cells. Many microorganisms adopt filamentous shapes caused by cell division arrest, concomitant with the continuous cell volume growth, in response to a variety of stressful environments, including treatment with  antibiotics. The presence of filamentous cells was assessed on Acinetobacter species in the presence of environmental stressors, and a description of A. baumannii ATCC 19606T “worm-like” cells has been reported, but their lifecycle and antibiotic susceptibility has not been investigated by the time of our experiments.  In our work cellular elongation in A. baumannii cells has been observed in both logarithmic and stationary phases by means of RCM imaging of SYTO9 and KK114 labelled cells. We consider this observation to be important as in many other bacterial species, the filamentation process is associated with antibiotic tolerance, but, to the best of our knowledge, this effect has not been deepened by the time of our experiments in A. baumannii. Although the resolving power of RCM is not as high as in the case of other optical nanoscopy techniques, the advantage it brings over conventional methods, such as CLSM, was found to be sufficient for identifying direct evidence of the filamentation process. The results of this study are presented in detail in an article published in IEEE Journal of Selected Topics in Quantum Electronics.

RCM images of the DNA distribution inside filamentous A. baumannii cells in the stationary phase (SYTO9 staining). Scale bars: 4 µm. A) DNA is homogenously distributed inside some elongated cellular elements (white arrow), or it is spatially organized in thickenings separated by constrictions (red arrows). B) Profile lines along two ROIs demonstrate two features: graph B1 (orange) DNA is distributed at the cell periphery; graph B2 (cyan) the fluorescence emission in elongated cells is lower compared to the case of coccoids.

Another line of study was to assess whether both label-free fixed or stained fixed tissues are equally useful for collecting NLO datasets in the purpose of subsequent automated classification by deep learning. In a study that was partially supported by GastroDeep, and published in Biomedical Optics Express, we evaluated the influence of the hematoxylin and eosin staining on the extraction of quantitative image parameters that describe the collagen structure, specific to image datasets obtained by polarization-resolved SHG microscopy. Thin serial tissue sections prepared as per standard histology protocol and stained with hematoxylin and eosin as well as their unstained pairs from skin and breast tissue samples were imaged. A fitting algorithm which provides ratios of the second order susceptibility tensor elements, fitting efficiency and average pixel intensities were assessed. While in the case of the susceptibility tensor elements ratios the change was inconclusive for the two tissue types which were under study, both the fitting efficiency and the average pixel intensity had a significant decrease for H&E-stained tissues, especially on the backward SHG pathway. By computing the forward-to-backward SHG ratio the influence of H&E staining on the collagen was connected to a possible modification in the SH-generating collagen shell. Our results show thus that similar quantitative analysis workflows applied to PSHG images collected on stained and unstained tissues yield different results, which can hinder the diagnostic accuracy if applied in an unsupervised manner. These findings suggest the need for novel PSHG image analysis methods and workflows that are specifically dedicate to stained or unstained tissue, and of methods that are capable to generalize better than current ones. Although our current results provide interesting findings on the influence of the H&E staining on the extraction of quantitative image parameters that describe the collagen structure, future experiments might shed more light on aspects which were currently inconclusive. Such future experiments might consider different collagen models (e.g., the generic model), different tissue types, or even different laser wavelengths.

Schematic representation of the imaging and quantification protocol: (a) images of two histology slides containing serial tissue sections, two stained with H&E (upper slide), and two unstained (lower slide, with the arrows indicating the position of the tissue sections); (b) large image depicting one H&E-stained tissue section; (c) image sets acquired on slides containing breast tissue; (d) polarization angle vs. SHG intensity (color-coded, with frames from 0 to 10 corresponding to polarization angles from 0° to 180° in steps of 20°) and images obtained from the FSHG image set and corresponding histograms; (e) image sets acquired on slides containing epithelial tissue. For (c) and (e) the MPM images are pseudo-colored: blue-color for FSHG, green-color for BSHG and red-color for autofluorescent tissue regions (probed by TPEF).

Another research track of GastroDeep refered to developing novel automated methods based on artificial intelligence capable to accurately distinguish pathological features in the collagen architecture of tissues. Various strategies have been tested. A successful approach identified is the joint use of popular pre-trained models, such as InceptionV3, ResNet50, with light custom-models custom developed taking into account the specific of images collected with non-linear optical microscopies (requiring no pre-training). Importantly, we observed that the combined use of such complementary models boosts the overall classification performance in the case of differentiating healthy from pathological collagenous tissues.

Schematic representation of the custom-developed FLIMBA model. Each convolutional layer (light yellow) is followed by a dropout layer with 0.2 probability (light green), followed by a 2×2 max-pooling layer. After the last max-pooling layer, a GAP and two FC layers (light purple), are added, each followed by a dropout layer with 0.5 probability (green). Last, an output layer with two softmax neurons is used.

The project work on Deep Learning classification of NLO images collected on Gastric Cancer tissues was successfully completed, an article presenting interim results is currently under preparation and will be submitted for peer-review in spring 2023.