Background
Among various imaging methods including magnetic resonance imaging, ultrasound, and computerized topography, contrast improvement is the ultimate goal for all medical imaging. In terms of the widely used medical white light imaging (WLI) (e.g., various cavity mirror and endoscopic), although it provides a higher resolution and faster imaging speed than MRI and CT to help medical doctors to discern changes of finer lesions and to better reduce or eliminate the blurring effect caused by heart bumping or pulmonary motion of the patient, however, its contrast is quite limited. With the conventional WLI, various complicated techniques including spatial registration, segmentation, lesion detection, and classification were required to process the image and differentiate the tissue into different clinical meaningful types [
1]. For example, during the global war to fight against the cervical carcinoma, which is the second most common gynecological malignancy in the world [
2], it is found that with the contrast provided by WLI, it is hard to build an accurate computer-aided diagnosis method to automate the cervical cancer screening process. While a CAD cervical cancer screening method is expected to improve the coverage of regular cervical cancer screening program and greatly prevent the occurrence of cervical cancer in underdeveloped countries and low-income regions [
2]. Accordingly, considerable efforts have been dedicated in the last two decades to automate the analysis of white light colposcopic images to support the medical decision process and to provide a data-driven channel for communications of findings [
1]. Recently, Intel and Mobile ODT organized a competition for the automatic analysis of digital colposcopies [
1]. Fernandes and colleagues gave a comprehensive review and summary of existing computer-aided diagnostic (CAD) method applicable to images provided by a conventional white light digital colonoscopy. They concluded that the existing WLI methodologies require a significant amount of manual labeling, including spatial localization of the lesions at an image level [
1]. To deepen the degree of automation of the cervical cancer screening process or any other lesion diagnosis process involving optical imaging, one important direction of effort is to enhance the contrast at the image acquisition stage, i.e., at the signal acquiring stage instead of at only the signal processing stage.
Accordingly, multiple authors suggest that spectral imaging method which combines the advantages of both digital image and spectral analysis may enhance the contrast between normal and abnormal tissues [
3,
4]. Spectral imaging may simplify the CAD process and reduce the amount of expert manual labeling. The fundamental principle of spectral imaging is that different chemical and biological molecules have a different, unique, and wavelength-dependent way to reflect, scatter, or absorb light. Accordingly, it is possible to use spectral difference to differentiate diseased regions, or to make optical biopsy. With the spectral difference-enhanced optical biopsy, a CAD method can be established for automatic and objective diagnosis of the precancerous lesions. Sterenborg proposed a spectral imaging method for diagnosing cervical tissues using an optical fiber to collect reflective lights from cervical squamous and achieved 89 and 80% of sensitivity and specificity, respectively [
5]. Recently Wang and coworkers proposed a multi-scale hyperspectral imaging method to detect cervical neoplasia at both tissue and cellular levels [
6]. The advantages of various existing spectral imaging based optical biopsy methods are non-invasive, objective, and has the potential to reduce the number of unnecessary biopsies. However, existing spectral imaging method obtains data by sequential scanning in the spatial domain to cover the area to be diagnosis. Therefore, it is time consuming to acquire the diagnosis needed data—it takes a while for existing spectral or hyperspectral imaging methods to scan the whole area of the cervix. Other limitations include the need for spatial registration, non-uniform illumination, high-cost and bulky setup, and sophisticated image processing [
7]. Instead of using spectral or hyperspectral imaging, it was found that as few as only three specific spectral bands are sufficient to classify the cervical tissues into normal, inflammation, and high-grade lesions [
7]. This is consistent with Benavides’s report that reflectance MSI at only a couple of wavelengths is needed to differentiate cervical cancer from its background [
8]. Multispectral reflectance imaging is highly preferable over hyperspectral imaging as it is more efficient, more compact, and cost effective [
7]. This is because unlike spectral or hyperspectral imaging, MSI does not spend time or effort to collect images of unwanted spectral bands. To further explore MSI, some researchers combine reflectance MSI with autofluorescence to investigate the combined value of both. Ren and coworkers proposed to combine multispectral reflectance, autofluorescence, and RGB imaging for noninvasive characterization of cervical intraepithelial neoplasia (CIN) [
9]. Their preliminary clinical results indicate that the added value of autofluorescence to the reflectance MSI is arguable considering the economic cost of the device and the additional operation in a clinical setting. So far, there is a trend to capture a couple of reflective spectral images to enhance the visual contrast in order to help the lesion-associated characteristic features stand out. For example, Olympus’s narrow-band imaging (NBI) method uses only two narrow bands corresponding to absorption characteristics of hemoglobin to enhance contrast to ensure that blood vessels in the mucosa stand out clearly [
10]. NBI method is successfully used for diagnosing bladder cancer [
11,
12], early gastric cancer [
13], pulmonary diseases [
14,
15], colon cancer [
16,
17], and cervical adenocarcinoma [
3,
11]. However, in most existing reflectance MSI including the Olympus NBI, the spectral images are sequentially acquired. They all involve scanning in the temporal domain. Hence, the spatial dislocation between images is inevitable, and spatial registration algorithm is required, which increases the computational load and reduces the efficiency.
Driven by the direct clinical needs of enhancing visual contrast for detecting early staged cancers at their early stages with a low-cost, competent, objective, convenient method, in this paper, we demonstrate a semi-automatic miniaturized snapshot narrow-band imaging (SNBI) method. Although the objective of SNBI is to enhance image contrast and to provide objective diagnosis, in this paper, it is evaluated for the tasks of cervical cancer screening. This is because it is of crucial importance to have a fast and automatic cervical cancer screening method to enlarge the coverage of regular screening population in the low-income and developing countries [
18]. Our method is specially designed to instantly capture four characteristic spectral images that maximize the difference between normal and diseased tissues. Using our SNBI approach, four spectral images are captured with zero-time-lag in between and through a common optical path, hence they are spatially co-registered and can be readily fused into a combined image with enhanced contrast between normal and abnormal tissues. Our goal is to make the refresh rate of the contrast-enhanced fused image as high as a conventional white light colposcopy. With the fused image, a simple semi-automatic CAD method can be developed to classify the cervical tissue into different and clinical meaningful grades of tissue types. The critical objective of our method is to automatically detect lesions, grade them, and delineate their boundaries with minor supervision, and to do all of these at fast speed of a video-refresh rate. This is highly desirable for its future application, especially on a large clinical scale, in cervical cancer screening [
1]. The advantages of the SNBI method include objective, invasive, chemical-free, instant results in vivo, and cost effectiveness. It is an ideal video camera to be integrated within a colposcopy to enlarge the coverage population of a cervical screening program in low-income countries and rural regions where medical resources are relatively scarce. The “
Methods and Experiment” section 2 describes the SNBI method and validation experiment. The “
Experimental Results” section presents the experimental results which indicate that the SNBI method in deeds enhance visual contrast and is capable to provide fast and objective diagnosis with weak supervision. The “
Discussion” section discussed the advantages of the SNBI and its significance to the cervical cancer screening. A brief conclusion is disclosed in the “
Conclusions” section.
Discussion
This paper proposed a snapshot NBI method with a semi-automatic CAD, i.e., a miniatured micro-arrayed SNBI method for automatic diagnosis. This SNBI method could instantly capture multiple images of cervix tissues at four characteristic bands centered at wavelengths of 415, 450, 525, and 620 nm. Since the four spectral images were spatially co-registered, they could be readily fused into a combined image with enhanced contrast between normal and abnormal tissues. With the fused image, a simple computer–aided diagnosis method is sufficient to classify the biological tissue into different and clinical meaningful grades of tissue types. This diagnosis method using Euclidean distance for classification, and it needed few doctor’s diagnoses as inputs. The preliminary results of the clinical evaluation experiment indicated that the method, without the aid of iodine and acetic acid reagents, could effectively classify the cervical tissue into meaningful pathological types of tissues. The classification is in good accordance with the common pathological diagnosis that acting as the current gold standard of clinical cervical cancer screening program.
The proposed snapshot NBI method with a semi-automatic CAD could achieve a highly efficient diagnosis at a refreshing rate equivalent to that of the underlying SCMOS monochrome camera, which is comparable to a conventional digital colposcopy. This is possible due to two reasons. First, the refresh rate of the generated fused spectral images is as high as the frame rate of the underlying SCMOS monochrome camera, as the multiple spectral images are captured with zero-time-lag-in-between at a single exposure. Second, only simple arithmetic operations are required to fuse multiple spectral images and the Euclidean distance classification algorithm contains only simple operations such as summation and comparison which can also be done in real time. The proposed SNBI method can update diagnosis result over 11 fps on a Pentium 1.6-GHz laptop. It is reasonable to conclude that the goal of achieving real-time video rate diagnosis by the proposed snapshot NBI method is proved to be feasible.
One interesting feature of the proposed SNBI method is its capability to accurately classify challenging regions where different medical experts may find it is hard to accept each other’s opinion. Further, the graphical layout of the quantitative classification results provides a convenient communication channel for the medical experts to discuss upon. Another advantage of the proposed SNBI method is its capability of detecting the boundary contours of different tissue types (Fig.
10). This is desirable for the clinics to know the exact location, size, and boundary of the diseased region, especially during image-guided sample collection procedure and precise surgical treatment.
Currently, we are integrating the SNBI video camera with an auto-focusing optical system with adjustable optical zoom from × 2 to × 26. After this, we may conduct an in vivo preclinical study using the SNBI imaging system to formally validate its diagnostic value for cervical cancer screening.
Conclusions
The proposed snapshot narrow-band imaging (SNBI) method enhances contrast and improves computer classification accuracy over conventional white light imaging (WLI) methods. The boundary contour between health tissue, cervical precancerous regions, and carcinoma in situ can be automatically delineated in SNBI. The SNBI contrast–enhanced method could make objective diagnostic result instantly and invasively with weak supervision where the expert identifies the presence of only a couple of lesions without explicating their boundaries. It generates automatic diagnostic results with clear boundary contours at over 11 fps on a Pentium 1.6-GHz laptop.
Further, it is miniature in size, cost-effective, and convenient to operate without the need of a sequence of data acquisition operations. Hence, the proposed SNBI is of great significance to enlarge worldwide the coverage of regular cervical screening program especially in low-income countries, and to live guide surgeries such as biopsy sample collection and accurate cervical cancer treatment. Although validated only with cervical cancerous tissues, in principle SNBI should work with any hemoglobin rich/scarce lesion detections.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.