Rapid hyperspectral photothermal mid-infrared spectroscopic imaging from sparse data for gynecologic cancer tissue subtyping

Abstract

Ovarian cancer detection has traditionally relied on a multi-step process that includes biopsy, tissue staining, and morphological analysis by experienced pathologists. While widely practiced, this conventional approach suffers from several drawbacks: it is qualitative, time-intensive, and heavily dependent on the quality of staining. Mid-infrared (MIR) hyperspectral photothermal imaging is a label-free, biochemically quantitative technology that, when combined with machine learning algorithms, can eliminate the need for staining and provide quantitative results comparable to traditional histology. However, this technology is slow. This work presents a novel approach to MIR photothermal imaging that enhances its speed by an order of magnitude. Our method significantly accelerates data collection by capturing a combination of highresolution and interleaved, lower-resolution infrared band images and applying computational techniques for data interpolation. We effectively minimize data collection requirements by leveraging sparse data acquisition and employing curvelet-based reconstruction algorithms. This approach enhances imaging speed without compromising image quality and ensures robust tissue segmentation. This method resolves the longstanding trade-off between imaging resolution and data collection speed, enabling the reconstruction of high-quality, high-resolution images from undersampled datasets and achieving a 10X improvement in data acquisition time. We assessed the performance of our sparse imaging methodology using a variety of quantitative metrics, including mean squared error (MSE), structural similarity index (SSIM), and tissue subtype classification accuracies, employing both random forest and convolutional neural network (CNN) models, accompanied by Receiver Operating Characteristic (ROC) curves. Our statistically robust analysis, based on data from 100 ovarian cancer patient samples and over 65 million data points, demonstrates the method's capability to produce superior image quality and accurately distinguish between different gynecological tissue types with segmentation accuracy exceeding 95%. Our work demonstrates the feasibility of integrating rapid MIR hyperspectral photothermal imaging with machine learning in enhancing ovarian cancer tissue characterization, paving the way for quantitative, label-free, automated histopathology. It represents a significant leap forward from traditional histopathological methods, offering profound implications for cancer diagnostics and treatment decision-making.

Figure 1

Schematic illustration of the O-PTIR optical configuration showing both the IR and green (532 nm) laser paths. Pulsed QCL at point (a) causes a photothermal expansion in the sample. A Continuous Wave (CW) green laser, indicated by (b), is collinearly directed onto the sample to serve as a probe beam. A dichroic mirror (c) merges the green and QCL beams, focusing them onto the sample (e) through a reflective Cassegrain objective (d). The resulting modulation in the intensity of the green light (f), scattered back from the sample, facilitates the measurement of its IR absorbance.

Figure 2

Comparison of High-Definition FT-IR and O-PTIR images of a cancerous Core. This figure illustrates the significant advantage of O-PTIR over FT-IR, showcasing its ability to overcome the diffraction limit, which results in enhanced spatial resolution. The improved image quality of O-PTIR is evident.

Figure 3

Microarray of ovarian cancer cores imaged by O-PTIR at the 1660 cm⁻¹ band. The data encompasses samples from 100 ovarian cancer patients. Variations in tissue biochemistry are highlighted by the color differences, demonstrating the rich biochemical information at the 1660 cm⁻¹ band, chosen for high-resolution reconstruction due to its significance in the fingerprint region. Scale bar: 1.5 mm.

Figure 4

Schematic for the data reconstruction algorithm used to enhance data acquisition speed. The figure illustrates how rectangular pixel-spaced data (0.5X5) acquired from the O-PTIR system is used to reconstruct 27 high-resolution, diffraction-limited band images. This method increases the data acquisition speed by 10X, yielding high-resolution images that offer more detailed information for improved segmentation of different cell types. The algorithm fuses spatial features from a high-resolution Amide I image with the linearly interpolated rectangular image via curvelet transform. This fusion preserves the biochemical information of each band image while accurately translating the spatial features of biological samples.

Figure 5

Method for interpolating low-resolution band images. We begin by performing a Fourier transform, followed by padding zeros along the Y-axis, and then applying a Gaussian filter to isolate lower frequencies. The interpolated image is obtained by taking the inverse Fourier transform.

Figure 6

Data fusion through curvelet transform involves taking the curvelet transform of both low-resolution band images and the high-resolution Amide I image. We obtain the high-frequency coefficients from the high-resolution image to achieve sharp edges, while low-frequency coefficients are obtained from the low-resolution band image. By combining these two and taking the inverse curvelet transform, we obtain the reconstructed band image.

Figure 7

Comparison of (a) interpolated, (b) computationally reconstructed, and (c) experimentally obtained high-resolution images. The comparison reveals that data collected at higher speeds with lower resolution can be effectively compensated for by our image sharpening method, which significantly improves upon the interpolated image.

Figure 8

Reconstruction accuracy vs pixel spacing. (a) Mean square error and (b) structural similarity (SSIM) averages vs pixel spacing. Data were collected from four cores at varying spacings along the Y-axis (0.5 μm×0.5 μm, 0.5 μm×1 μm, 0.5 μm×2 μm, 0.5 μm×3 μm, 0.5 μm×5 μm, 0.5 μm×10 μm, 0.5 μm×20 μm). The 0.5 μm×0.5 μm spacing image served as a reference. We calculated MSE and SSIM for the various spacings and reported the mean and standard deviation for the cores.

Figure 9

Comparison of stained image (a) identified as ground truth by a pathologist, with classifications by RF (b) and CNN (c) for two ovarian tissue cores. The RF model demonstrates significant improvement over our previously published results, attributed to the increased number of bands. Conversely, the CNN model achieves classification comparable to that of a pathologist’s analysis on a stained tissue microArray (TMA), owing to its utilization of both spectral and spatial information.

Figure 10

ROC curves and associated AUC values for each tissue subtype. CNN (blue line) demonstrates superior results compared to RF (dashed orange line) across all tissue subtypes: (a) epithelium, (b) necrosis, and (c) stroma.

Figure 11

Classification results of 100 cores with (a) RF and (b) CNN. Red, Green, and Blue channels correspond to epithelium, stroma, and necrosis respectively. The scale bar is 1.5 mm