Label-free multiphoton excitation imaging as a promising diagnostic tool for breast cancer

2.1 Clinical specimens

Breast tissue samples consisting of 30 carcinoma tissues and 25 normal tissues were collected postoperatively from 32 patients with breast carcinoma from the Cancer Institute Hospital of the Japanese Foundation for Cancer Research (Table 1). Samples were collected from the center of the tumor or the residual normal breast tissue surrounding the tumors. They were immediately immersed in phosphate-buffered saline with 10% fetal bovine serum and penicillin/streptomycin and delivered to the imaging room at Osaka University. Immediately following imaging, the samples were fixed in 10% neutral buffered formalin (Muto Pure Chemicals) and processed routinely for paraffin embedding. All patients were histologically diagnosed preoperatively using biopsies. All patients provided written informed consent in accordance with the ethics committee requirements of the hospital and the Declaration of Helsinki. This study was conducted with the approval of the ethics boards of the Cancer Institute Hospital of the Japanese Foundation for Cancer Research and the Osaka University Graduate School of Medicine. The Osaka University Graduate School of Medicine Institutional Review Board approved the study protocol on December 17, 2015 (No. 15369). The Cancer Institute Hospital of the Japanese Foundation for Cancer Research Institutional Review Board approved the study protocol on August 18, 2016 (No. 2016-1059).

2.2 Imaging of human breast tissues and group separation of data files

The imaging system consisted of an upright microscope (A1RMP+; Nikon) driven by a laser (Chameleon Vision II Ti: Sapphire; Coherent, Inc.) tuned to 780 nm, and an upright microscope equipped with a × 25 water immersion objective (CFI75 Apo 25 × W MP/NA 1.10; Nikon). Tissue samples were positioned with the observation surface facing upward and overlaid with a coverslip to place a drop of water between the sample and the objective lens. To detect multiphoton excited fluorescence and second harmonic generation (SHG) emission signals, we used 390/18-nm, 480/40-nm, and 629/56-nm band-pass filters. To acquire a series of image files from one sample, we collected image stacks at 3-μm vertical steps at a depth > 30 μm below the sample surface with 1.0× zoom and 1024 × 1024 X-Y resolution (0.50 μm per pixel). Following imaging, we calibrated the z-coordinate (orthogonal to the observation plane) of the stacked image files by defining z = 0 as the first height at which the fluorescent signal was detected in almost the entire image area. Moreover, images of the surface side with z ≤ 30 μm were used in this study.

In normal tissues, the adipose tissues, fibrotic stromal tissues, and vascular structures were observed in addition to the ductal-lobular structures. In cancerous tissues, we could recognize invasive breast carcinoma, mucinous carcinoma, and ductal carcinoma in situ in addition to the surrounding normal tissue components. An experienced pathologist linked the ground truth to each image file. We divided all image files from normal and cancerous tissues into two groups, depending on the case. Image data of group A (1765 images from 12 patients; consisting of 784 images from normal tissues, 91 images without malignant findings from cancerous tissues, and 890 images with malignant findings from cancerous tissues) were used for the construction of classification models, including training and validation. We used the training data set for the construction of the learning model candidates, and the validation data set was used to select the learning model with a high generalization performance. Lastly, we analyzed the classification performance of the determined learning model using the files of group B (2393 images from 20 patients, consisting of 899 images from normal tissues, 138 images without malignant findings from cancerous tissues, and 1356 images with malignant findings from cancerous tissues).

In the spectral analyses, we used the spectral detector of the A1RMP+ microscope (Nikon). Tissues were imaged to record their emission spectra from 380 nm to 630 nm (collected in 25 bins, each approximately 10 nm wide) under 780 nm excitation. Each fluorescence intensity was recorded at 4096 gray intensity levels (12 bits).

2.3 Construction of image classification model

All image files, in the form of squares with a side of 1024 pixels, were normalized before the analysis such that the mean value was 0 and the standard deviation was 1. Moreover, they were cropped into 64 squares with sides of 128 pixels (termed “image tiles”). After cropping, we separated the validation data set by randomly selecting 20% of the image tiles from group A; the resulting 90,368 and 22,592 tiles were prepared for training and validation, respectively. Subsequently, the image tiles with a signal-to-noise ratio < 20 and an average brightness value <0.10 were defined as blank. The signal-to-noise ratio was based on a pseudo-noiseless image, which was constructed with a median filter. Based on these criteria, 9908 tiles in the training set, 2498 tiles in the validation set, and 12,258 tiles in the test set (group B) were judged as blank. After separating the blank image tiles, we established a classification model for the nonblank image tiles using Inception-v3, a type of deep-learning architecture. The code was implemented in Python using TensorFlow 2.0, and Keras 2.3, which are open-source software libraries for machine intelligence. The Adam optimization algorithm was used with an epoch of 100, a learning rate of 0.020 and a batch size of 128. A total of 80,460 and 20,094 image tiles from group A were used for training and validation during training, respectively. Following the construction, all 140,894 nonblank image tiles from group B were tested for an automatic classification of the presence of cancer cells. After the test, all image tiles from group B (153,512 tiles), involving those judged as blank (12,258 tiles), were evaluated to calculate detection accuracy. The image tiles judged as representing cancer were classified as “positive.” In contrast, image tiles judged as showing no cancer cells as well as the blank ones were classified as “negative.”

2.4 Histology

Paraffin-embedded specimens were cut into 4-μm-thick sections and stained with hematoxylin and eosin (H&E) using a standard protocol.

2.5 Statistical analyses

For comparisons between groups with non-Gaussian distributions, we performed the Mann–Whitney U test to calculate P values. Accuracy represents the number of correctly classified observations over the total analyzed data number. Precision represents the ratio of correctly predicted positive observations to the total predicted positive observations. Recall, a synonym for sensitivity, represents the ratio of correctly predicted positive observations to all observations in actual positivity. F-measure represents the weighted average of precision and recall. To compare diagnosis accuracies by experienced pathologists between conventional histopathological method using HE staining and MPM images, Cohen's kappa coefficient was calculated as a measure of interinspection reproducibility.

3.1 Autofluorescent profiles of human fresh breast tissues

We analyzed the autofluorescent spectral profile of fresh human breast tissues (Figure 1A). Previous studies using human fresh tissues other than breast reported on detecting sufficient autofluorescent signals by an excitation wavelength ≤800 nm.^{12, 13} Therefore, we employed an excitation wavelength of 780 nm in the current study. We continuously acquired emission signals between 380 nm and 630 nm under the aforementioned conditions (Figure 1B and Figures S1–S3). Epithelial cells in the normal mammary duct displayed an emission peak at 480 nm (red region of interest [ROI] in Figure 1B). Similar autofluorescence was observed in adipose cells around the mammary gland (purple ROI in Figure 1B) and fibrous structures (blue and black ROIs in Figure 1B). We could also detect a dot-like autofluorescence near the edge of normal ducts with an emission peak at 510 nm (cyan ROI in Figure 1B). In addition, we detected a sharp and narrow emission fluorescence peak at 390 nm from the fibrous structures (blue and black ROIs in Figure 1B). This represented half the excitation wavelength, thereby indicating that we observed fluorescent signals by SHG derived from fibrillar collagens.¹⁴ In cancerous tissues, the emission peak of cancer cells was similar to that of normal epithelial cells (orange ROI in Figure 1B). Contrarily, immune cells around the cancer cells displayed autofluorescence with an emission peak at 570 nm (magenta ROI in Figure 1B). Thin fibrous structures near the cancer cells emitted less autofluorescence, except for SHG (green ROI in Figure 1B). Therefore, human breast tissue emitted multiple autofluorescent signals via near-infrared ray excitation. It principally consisted of three major peaks, which prompted us to describe the histological features in detail with these fluorescent signals.

3.2 MPM images of breast tissue enabled pathologists to estimate normal or cancerous lesions accurately

Based on the abovementioned spectral analysis, we established an MPM imaging system using three channels with different band-pass filters (Figure 2A). Next, we performed label-free MPM imaging of fresh unstained breast tissues, including normal and cancerous lesions. H&E-stained sections from identical tissues used for MPM imaging were used to confirm the diagnosis. In fat tissues around the mammary gland, we identified adipose cells with circular morphology by detecting autofluorescent signals (first row of Figure 2B). Epithelial cells were lined neatly on the basal side of the lumen in mammary ducts in normal tissues. Cellular nuclei were identified as signal-void regions and did not display obvious enlargement. In addition, dot-like autofluorescence was detected along the edge of the ductal structures (second row of Figure 2B). In contrast, tumor cells of ductal carcinoma in situ revealed irregular proliferation in the lumen with nuclei enlargement on MPM imaging. Moreover, the dot-like autofluorescence at the lesion edge was unclear (third row of Figure 2B). In invasive carcinoma tissues, the tumor cells proliferated irregularly without luminal formation (fourth row of Figure 2B). These histological features recognized in MPM images resembled those of conventional H&E-stained images (fourth column of Figure 2B). Multiphoton excitation microscopy images of both normal tissues (from 25 patients) and cancerous lesions (from 30 patients) were randomly examined by two experienced pathologists, who could provide correct diagnoses using only the MPM images with a kappa coefficient of 1.0 (Table 2). These results demonstrated that experienced pathologists can estimate normal or cancerous lesions using only our MPM images with adequate high accuracy.

3.3 Image classification with AI algorithm enabled to detect cancer lesion in small area without pathologist's intervention

When inspecting a small lesion or tumor margin, cancer cells are usually found only in a small area (if any). Therefore, we constructed an AI classification algorithm that could detect cancer cells that exist in a portion of the captured image (Figure 3A). We randomly divided all image files into two groups, depending on the patients. Subsequently, we cropped all files into 64 squares with 64 μm side each, or “image tiles,” and a pathologist linked ground truth to each tile. Considering the inclusion of blank image tiles in the edge area of the original images, we separated them based on the signal-to-noise ratio and average brightness value. Nonblank tiles from group A were used to construct an image classification model based on Inception-v3, a type of CNN architecture. We used image tiles from group B to examine the classification algorithm. Both image tiles judged as involving no cancer cells as well as blank ones were classified as “negative.” In contrast, those judged as involving cancer were classified as “positive.” The utility test of our classification algorithm, performed with 153,152 tiles from group B, indicated accuracy, precision, recall, and F-measure of 0.934, 0.827, 0.942, and 0.881, respectively (Figure 3B and Table 3). Moreover, to evaluate the detection accuracy of the malignant area in the original captured image file (consisting of 64 tiles), we defined the index termed “malignant probability” as the ratio of positive tiles in an original captured image. This index was calculated for each analyzed image file from group B, which revealed a significant difference between the normal and cancerous images (Figure 3C). Under the threshold of 0.0625 (from the receiver operating characteristic curve in Figure 3D), we classified 1314 of 1356 cancerous files, and 974 of 1037 normal files were classified as “malignant” and “nonmalignant,” respectively, with a sensitivity, specificity, and area under the curve of 0.969, 0.939, and 0.989, respectively (Figure 3D). These results indicated that a combination of label-free MPM imaging and AI-based image classification enabled the detection of cancer cells in a small area.