Journal of Biomedical Optics最新文献

Significance: Single-chip multispectral imaging sensors with vertically stacked photodiodes and pixelated spectral filters enable advanced, real-time visualization for image-guided cancer surgery. However, their design inherently reduces spatial resolution. We present a convolutional neural network (CNN)-transformer demosaicing algorithm, validated on both clinical and preclinical datasets that effectively doubles spatial resolution and improves image quality-substantially enhancing intraoperative cancer visualization.

Aim: We present a CNN-transformer-based demosaicing approach specifically optimized for reconstructing high-resolution color and NIR images acquired by a hexachromatic imaging sensor.

Approach: A hybrid CNN-transformer demosaicing model was developed and trained on color-image datasets, then rigorously evaluated on color and NIR images to demonstrate superior reconstruction quality compared with conventional bilinear interpolation and residual CNN methods.

Results: Our CNN-transformer demosaicing method achieves an average mean squared error (MSE) reduction of $\sim 85\%$ for color images and 76% for NIR images and improves structural dissimilarity by roughly 72% and 79%, respectively, compared with state-of-the-art CNN-based demosaicing algorithms in preclinical datasets. In clinical datasets, our approach similarly demonstrates significant reductions in MSE and structural dissimilarity, substantially outperforming existing CNN-based methods, particularly in reconstructing high-frequency image details.

Conclusions: We demonstrate improvements in spatial resolution and image fidelity for color and NIR images obtained from hexachromatic imaging sensors, achieved by integrating convolutional neural networks with transformer architectures. Given recent advances in GPU computing, our CNN-transformer approach offers a practical, real-time solution for enhanced multispectral imaging during cancer surgery.

Significance: Leukemia, a complex hematological cancer, poses significant diagnostic challenges due to the heterogeneity of leukemic cells, inter-observer variability, and lack of standardized analysis methodology. Accurate and rapid cell classification is essential to improve clinical management, optimize treatment, and reduce diagnostic errors.

Aim: We propose an innovative approach combining optical scanning holography (OSH) and active contour (AC) models to automate the classification and segmentation of leukemic cells with increased accuracy.

Approach: OSH is used to capture the phase current of leukocytes, providing a cost-effective, noninvasive, and simplified alternative to conventional techniques. AC models are used to improve cell segmentation. Analysis of the maximum amplitude values of the phase current allows rapid and fully automated classification.

Results: The proposed approach shows a significant improvement in terms of reliability, speed, and reproducibility compared with existing methods. The integration of OSH and AC enables robust segmentation and efficient classification of leukemic cells.

Conclusion: This method provides a reliable, rapid, and systematic solution for the accurate diagnosis of leukemia, enabling optimized therapeutic management.

Significance: Accurate cell classification is essential in disease diagnosis and drug screening. Three-dimensional (3D) voxel models derived from holographic tomography effectively capture the internal structural features of cells, enhancing classification accuracy. However, their high dimensionality leads to significant increases in data volume, computational complexity, processing time, and hardware costs, which limit their practical applicability.

Aim: We aim to develop an efficient and accurate cell classification method using 3D refractive index (RI) point cloud data obtained from holographic tomography, focusing on reducing computational complexity without sacrificing classification performance.

Approach: We transformed 3D RI voxel data into point cloud representations using segmented equilibrium sampling to substantially decrease data volume while retaining crucial structural features. A deep learning model, named RI-PointNet++, was then specifically designed for RI point cloud data to enhance feature extraction and enable precise cell classification.

Results: In experiments classifying the viability of HeLa cells, the proposed method achieved a classification accuracy of 93.5%, significantly outperforming conventional two-dimensional models (87.0%). Furthermore, compared with traditional 3D voxel-based models, our method reduced computational complexity by over 99%, with floating-point operations of only 1.49 G, thus enabling efficient performance even on central processing unit (CPU) hardware.

Conclusions: Our proposed method provides an innovative, lightweight solution for 3D cell classification, highlighting the considerable potential of point cloud-based approaches in biomedical research applications.

Significance: Near-infrared spectroscopy (NIRS) imaging modalities are used to provide noncontact measurements of tissue oxygenation in diabetic foot ulcers. However, the curved surface of the diabetic foot introduces inaccurate tissue oxygenation measurement. The changes in spatial NIRS optical measurements may result from variations in the underlying physiology or from the curvature of the tissue surface. Therefore, the effect of tissue curvature must be accounted for to ensure the accurate measurement of tissue oxygenation (or hemoglobin parameters) in clinical applications.

Aim: Our aim is to develop and validate mathematical curvature correction models to account for the effects of tissue curvature on diffuse reflectance (DR) in NIRS imaging and assess their effect on the hemoglobin parameters as well.

Approach: Monte-Carlo-based light propagation simulations were performed to develop correction models and applied to three-layered curved geometries in MCMatlab. Four curvature correction models based on height and/or angle were developed via Monte Carlo simulation studies. All the correction models were applied to the simulated DR signals obtained from various curved geometries (concave, convex, and wound-mimicking) using Gaussian light sources at 690 and 830 nm. The effect of correction models on DR signals and hemoglobin parameters was determined.

Results: Simulation results showed that a concave curved surface did not require correction, whereas convex and wound-mimicking geometries showed a reduced median error upon using an empirical height/angle correction model. In addition, the correction model also reduced the median error significantly for the oxygen-saturation-based hemoglobin parameter in both the convex and wound-mimicking geometries.

Conclusions: The developed mathematical model effectively corrected tissue curvature effects in NIRS DR signals and hemoglobin parameters for wound-mimicking irregular geometry. Ongoing work focuses on experimental validation of these correction models on curved phantoms, prior to in vivo imaging studies.

Significance: Two-photon microscopy is widely used for in vivo imaging of vasculature in rodents and requires the labeling of blood plasma with fluorescent dyes such as indocyanine green (ICG). However, a major limitation of ICG is its rapid clearance from the body, which restricts its use in extended imaging sessions. We address and overcome that limitation, enabling longer in vivo imaging sessions.

Aim: We aim to investigate the feasibility of using liposomal nanoparticles that, when used to encapsulate ICG, significantly increase the circulation time of the vascular label in the rodent body.

Approach: We conducted in vivo imaging experiments with unencapsulated (free) ICG and liposomal ICG (L-ICG) and compared the retention of ICG in the vascular network over a duration of 75 min.

Results: In comparison to a retention time of around 20 min for free ICG, we find that liposomal encapsulation improves the vascular retention time of the dye to at least 75 min. The improvement in retention time using the encapsulation technique was consistent across imaging experiments conducted in five mice.

Conclusion: The rapid clearance of ICG from the rodent body can be overcome using liposomal encapsulation, making prolonged in vivo imaging feasible.

Significance: Mueller polarimetric imaging shows great promise for differentiating neoplastic from healthy brain tissue during neurosurgery. However, validating algorithmic approaches is limited by the scarcity of substantial tumor border zones in ex vivo samples, limiting comprehensive analysis of tumor margins.

Aim: We propose a protocol to build a database of histologically annotated polarimetric images from formalin-fixed whole-brain sections. We focus on validating the image alignment pipeline on healthy tissue.

Approach: To address the size mismatch between samples and the field of view of imaging instruments, we developed an automatic reconstruction pipeline to create large-scale polarimetric images from smaller raster-scanned tiles. Matching points between reference photographs and tile images allowed precise alignment. Similarly, fractionated histological sections were reconstructed and accurately aligned with the polarimetric data to serve as ground truth.

Results: The integrated reconstruction and alignment approach enabled large-scale, spatially co-registered polarimetric and histological imaging, supporting a more detailed investigation of tissue polarimetric parameters. The database thus created will facilitate the training and evaluation of segmentation models.

Conclusions: The developed method improved polarimetry-based brain tissue mapping by linking polarimetric parameters with histological features, enhancing the quality and quantity of data available for training and evaluating segmentation models. Although initially applied to brain tissue, the protocol could be extended to other organs to support broader studies of polarimetric tissue characterization.

Significance: We address the challenge of inadequate force feedback in laparoscopic surgery, which increases the risk of vessel injury. By integrating fiber Bragg grating (FBG) sensors with a laparoscopic grasper and employing a convolutional neural network combined with long short-term memory (CNN-LSTM) algorithm, this approach enables real-time, accurate vessel identification, potentially reducing surgical complications.

Aim: Laparoscopic surgery is often hindered by inadequate force feedback, especially in complex scenarios involving tumor invasion and pelvic-abdominal adhesion, leading to challenges in locating blood vessels and an increased risk of vessel injury. Thus, it is desirable to develop a laparoscopic system capable of distinguishing the location and type of the vessels during surgery, which requires a compact and highly sensitive sensor integrated with a laparoscopic grasper.

Approach: We present an innovative laparoscopic grasper integrated with FBG force sensors for real-time force feedback, employing silicone and porcine vessel models to simulate varying depths and tissue coverage. The device successfully captured specific vessel signals, which were processed through a CNN-LSTM algorithm, enabling real-time vessel identification in minimally invasive surgery (MIS).

Results: The intelligent laparoscopic grasper successfully obtained distinct vessel signals under varying conditions. As a result, the mean vessel gripping force for porcine vessel model III was 0.059 N under fatty tissue and 0.032 N under muscle tissue ( $p<0.001$ ). The CNN-LSTM algorithm achieved a precision of 97.06% in vessel identification across different tissue coverages.

Conclusions: The FBG sensor-integrated laparoscopic grasper, assisted by the processing of the CNN-LSTM algorithm, demonstrated the ability to identify vessels ex vivo across different models. This technology holds potential for real-time and accurate vessel identification during MIS, which could significantly reduce the occurrence of unnecessary vessel injuries.

Significance: Peritoneal dissemination is the major mechanism of how ovarian cancer (OC) spreads. It features tumor outgrowths in the form of multicellular spheroids, their detachment from the primary site, and their implantation in the peritoneal cavity. To understand this process, analyzing the outgrowths at their native locations within the female reproductive system is essential. However, in vivo study of the OC outgrowths remains unattainable primarily due to the lack of in vivo imaging approaches to probe such small tumor structures at a high resolution.

Aim: We address this technical challenge by establishing in vivo high-resolution 3D imaging of the OC outgrowths in the mouse model.

Approach: This in vivo imaging approach relies on optical coherence tomography (OCT) for 3D label-free imaging and an intravital window to bypass the mouse skin and muscle layers. To demonstrate the imaging capability, we use TgMISIIR-TAg transgenic mice that develop spontaneous epithelial OC. The normalized surface lengths of the ovary and the OC outgrowth are measured from OCT images to characterize the tissue morphology. Immunohistochemistry staining is employed to confirm the presence of transgene-positive cells in OC and outgrowths.

Results: We present the first in vivo high-resolution 3D image of the OC outgrowths in the mouse model. The tissue morphology and structure of OC outgrowths have striking differences from the normal ovary, which is quantitatively assessed and compared. We further show that OC outgrowths within and growing out of the ovarian bursa, revealing the difference in their surface morphologies. We also present the detached OC outgrowths and the fluid-filled chambers inside OC, both with 3D quantifications showing the heterogeneity of their volumes.

Conclusions: This in vivo OCT imaging approach in the mouse model enables high-resolution assessment of detailed 3D structures of OC outgrowths, paving the way for in vivo study of the OC dissemination process.

Significance: Melanoma's rising incidence demands automatable high-throughput approaches for early detection such as total body scanners, integrated with computer-aided diagnosis. High-quality input data is necessary to improve diagnostic accuracy and reliability.

Aim: This work aims to develop a high-resolution optical skin imaging module and the software for acquiring and processing raw image data into high-resolution dermoscopic images using a focus stacking approach. The obtained hyperfocus images should significantly enhance the diagnostic performance of total body scanners in clinical settings.

Approach: We employed focus stacking to merge multiple images, each with a limited depth of field, into a single hyperfocus image, ensuring every part of the skin is in clear focus. The method was implemented in the high-resolution imaging module using an electrically tunable liquid lens to quickly capture a series of differently focused images in vivo. Algorithms were developed for image alignment, focus measurement, and fusion, with the addition of deep learning-based super-resolution techniques to further enhance image quality. A classification model was trained to provide an artificial intelligence (AI)-based lesion classification.

Results: The developed optical imaging system successfully produced noncontact dermoscopic images with complete focus across all skin topographies. The hyperfocus images obtained demonstrated high resolution of $28\mu m$ and captured focus stacks at 50 frames per second, ensuring rapid acquisition and patient comfort, however, with some variance in resolution of individual lesions compared with contact-based dermoscopy standards.

Conclusions: The focus stacking-based approach for noncontact dermoscopy improves the quality of diagnostic images by ensuring an all-in-focus view of differently shaped skin lesions, essential for early melanoma detection. Although the approach marks a substantial improvement in noninvasive skin imaging, the use of super-resolution techniques requires careful consideration to avoid compromising the authenticity of the raw data. This work enables the usage of advanced imaging and AI techniques in total body scanners for early melanoma detection in clinical practice.

Significance: The spatial and temporal distribution of fluorophore fractions in biological and environmental systems contains valuable information about the interactions and dynamics of these systems. To access this information, fluorophore fractions are commonly determined by means of their fluorescence emission spectrum (ES) or lifetime (LT). Combining both dimensions in temporal-spectral multiplexed data enables more accurate fraction determination while requiring advanced and fast analysis methods to handle the increased data complexity and size.

Aim: We introduce two methods, a phasor and a feedforward neural network (FNN) analysis, to extract fluorophore fractions from temporal-spectral data. These methods aim to handle the increased data complexity and size of temporal-spectral multiplexed data and therefore enable access to a more accurate and fast fraction determination.

Approach: The phasor analysis determines the fraction in each dimension and combines them, whereas the FNN is trained using artificially mixed data. Both methods are compared with the reference method using linear combination-based curve fitting (FIT). The methods are tested in a two-component scenario of exogenous fluorophores with different ES and LT and in a three-component scenario of endogenous fluorophores with similar ES and different LT.

Results: In this case, the phasor analysis showed the lowest absolute errors in the fraction determination (1.4% two-component, 4.7% three-component), outperforming the FNN (6.3%) and FIT (8.7%) analysis, which are both not able to recognize all fluorophores in the three-component scenario. The computational effort was reduced by roughly a factor of 6 (Phasor/FNN) compared with FIT.

Conclusions: Both methods demonstrate substantial advantages over common fitting, offering a faster and more accurate determination of fluorophore fractions. These advancements make temporal-spectral multiplexed data more accessible and practical, particularly for high-speed applications.