Page 12 - ITPS-6-2

INNOSC Theranostics and
Pharmacological Sciences Theranostics in neurosurgery

Theranostics within the surgical resection of brain for automated analysis of processed images. Depending
tumors primarily emphasizes a means to supplement on the specific disorder and pathology, the capacity to
surgical resection with the intraoperative diagnosis train will vary. For example, pathologies with a higher
of tissue samples. Characteristics granting speed and prevalence of pathognomonic features will lend themselves
accuracy have been increasingly explored within the field, to greater ease in distinction . In a non-inferiority
[61]
paving the way for deep learning. Deep learning is a form randomized controlled trial, Hollon et al. assessed SRH
of machine learning and artificial intelligence that is giving paired DCNN to conventional H&E pathological analysis
way to the development of more diagnostic technology across 278 cases . Sister samples were generated and
[61]
within automated pathology for neurosurgical resections. designated to one arm of the trial. The overall diagnostic
Notably, a step within IOC pertains to the preparation accuracy was 94.6% for the DCNN arm and 93.9% for
of tissue samples, which lends itself to successful and the control, suggesting that accuracy is not sacrificed at
accurate diagnoses. Conventionally, tissue samples the expense of augmented diagnostic speed. While IOC
have been processed as frozen sections and subjected to may consume approximately 20 min in duration, DCNN
processing and labeling. Processing involves three steps to is implemented on a nearly real-time scale, as shown in
ensure the tissue is suitable for fixation within supportive Figure 3 [62-64] . Further, its potential has been illustrated in
molds. The steps are: (i) Dehydration with agents such as other studies that coupled this modality to other forms of
ethanol or isopropanol, among other alcohols; (ii) clearing, image preparation, namely, THG microscopy. Hence, the
involving agents, such as xylene; and (iii) infiltration with adoption of machine learning within IOC serves utility in
a medium of choice, such as paraffin [51,52] . The invasiveness the face of efficiency.
of these steps may carry downstream consequences for
the visualization and analysis of slide samples . These 4.2. THG microscopy
[53]
consequences can manifest as structural variations, As discussed, the processing stage plays a salient role
including tissue shrinkage, protein denaturation, the in the diagnostic ability provided by machine learning
resolution of macromolecules, and the intended degree and this subset of neurosurgical theranostics. Similar to
of staining [54,55] . Conventionally, the processed specimen SRH, THG is a multiphotonic technique that utilizes a
is labeled with hematoxylin and eosin (H&E) staining . label-free application. It uses three photons to produce a
[56]
More recently, two techniques have been explored which single photon with the sum of their energy in a process
include stimulated Raman histology (SRH) and third known as photoconversion . The produced images rely
[65]
harmonic generation (THG) microscopy [57,58] . on the inherent contrast of the visualized material to
4.1. SRH create high resolution, bypassing the detriments of certain
procedures, such as photobleaching or reactive oxygen
SRH, a modality that allows for the analysis of tissue species (ROS) production [66,67] . A study of 45 samples
through an unlabeled and unprocessed method for assessing THG efficacy in distinguishing gliomas from
preparation and staining, contributes to the greater non-tumorous tissue (n = 37 glioma and n = 8 normal)
preservation of important molecular structures, which was described by Blokker et al. According to the study,
[68]
supports downstream analyses. Developed in 2008, SRH THG has an imaging capture speed that is 8 times faster
is an infrared microscopical technique that utilizes the than SRH, yielding significant advantages regarding speed.
vibrational frequencies within the chemical bonds of The researchers applied fully convolutional networks
proteins, lipids, and DNA; it generates images highly (FCN), which is a form of deep learning, based on a set
reminiscent of H&E-produced imaging [59,60] . SRH has of image-level features determined by three pathologists.
footholds within spontaneous Raman scattering, which This was applied to assess the binary diagnostic ability to
is its original predecessor. This method, however, is distinguish between glioma and non-tumor. Overall, the
multiphotonic and utilizes two lasers to generate the accuracy was 79% with a mean average precision of 0.83.
desired emission signal through stimulated rather The results of this study in addition to those of the Hollon
than spontaneous excitation [59] . Overall, the potent et al. corroborate evidence for the diagnostic efficacy
[61]
variation within the hydrocarbon bonds of the sample of machine learning and artificial intelligence within
creates ample contrast for quicker image resolution and IOC while still granting a superior diagnostic speed.
generation. Nonetheless, development in this field remains relatively

The capabilities of machine learning in the form of nascent. The latter of the two studies solely assessed the
deep convolutional neural networks (DCNNs) have also diagnosis of gliomas without other intracranial tumors.
been explored in the context of SRH. DCNNs utilize The success is extrapolatable to other forms of tumors, such
trainable features based on histological patterns to allow as meningiomas and glioblastomas, based on the former

Volume 6 Issue 2 (2023) 6 https://doi.org/10.36922/itps.417