Quantitative lung morphology: semi-automated measurement of mean linear intercept

Environmental exposures due to ambient particulate matter (PM), tobacco products or other occupational exposures may cause morphologic changes in lung parenchyma. Methods to quantify these structural changes have been an area of investigation spanning several decades [1‐5]. These measures are important since they reflect clinically significant structural changes that can be a focus of therapeutic assessment [1, 2]. Mean linear intercept (MLI) is a measure of morphometric change based on serial measurements of the lung using test lines, but cannot be directly interpreted as a measure of alveolar size. MLI can be most directly interpreted as the mean free distance between gas exchange surfaces in the acinar airway complex [1]. Importantly, the use of MLI in estimation of the structural substrate of lung function is only applicable when considered a function of the total volume and surface area of the lung due to the lung’s structural complexity and complex inflational behavior [6]. Clinically, MLI has been associated with emphysematous changes to the lung and lung injury related to smoke and PM exposure [7, 8].

The degree of morphologic changes in affected lung may be accomplished using stratified random field selection and subsequent qualitative morphometric and pathologic analysis, such as the Destructive Index [3]; however, such methods may introduce unnecessary bias, so quantitative analyses such as MLI are preferred.

Importantly, there are several limitations in regards to MLI measures. There is no gold standard for the assessment of MLI [9]. MLI can be estimated by point or intersect-counting (superimposing a test line on a field and counting the number of times the line intercepts an alveolar septum, then using appropriate formulae) or directly measured by superimposing a test line on an image and measuring the distance along the line between consecutive intersections with alveolar septa [3]. Both intercept-counting and direct measurement, when performed manually, may take days to process even a few specimens. Additionally, there are no accepted values pertaining to sampling breadth, with the number of chords assessed per lung ranging from hundreds to thousands [3, 10‐12].

Some of these concerns can be addressed using custom-built software that can achieve fully automated measuring processes; however, these are often technically intensive and rife with potential error, including challenges related to the true identification of non-alveolar structures, and a tendency of automated methods to underestimate MLI compared to manual measurements, as reported by Sallon et al., as well as Langston and Thurlbeck [13, 14]. In addition, this software is not made freely available to the public; therefore, there is need for a free, open-source, semi-automated solution. Such a method would allow a degree of supervision to control for spontaneous error and produce reliable results. By adapting the system of test lines traditionally used in manual quantitative assessment for implementation in a semi-automated method, our objective is to produce reliable measures of MLI. We utilized a modified version of our semi-automated measure of MLI in our recent publication [8]. Furthermore, the method was utilized in several abstracts and presentations [15‐21]. Additionally, we developed a guide to determine the optimal number of chords for measurement to achieve accurate estimation of MLI per field of view. Our method relies on free, open-source software, does not require more than basic knowledge of computers, and, with the use of provided macros and plugin, has the potential to process hundreds of images in seconds (s) [22].

Overall, there is a significant need for a free method of lung histology quantification. MLI is a common measure used in pulmonary research; however, MLI measures have been difficult to obtain due to either the need for time consuming hand-measurements under a microscope via repeated superposition of a test line, or the need for proprietary software that may not be readily available. Despite availability of image-processing software, new computational tools remain under-exploited for application to the measurement of MLI. Our goal was to identify, optimize, and make readily available a method to measure MLI using freely available software. In addition, we wanted a method that would allow for an assessment of a significant amount of the lung.

Our method uses a novel combination of existing image processing functions that minimizes prerequisite, specialized computer-programming knowledge. While Fiji is used in this paper, any image processor with the ability to isolate test lines from background, overlay a semi-transparent layer on another image, and select non-contiguous pixels based on color, could also be used.

Processing one image (1000 pixels × 1000 pixels) in roughly 0.03 s, our method is faster than pre-existing methods. If performed manually, these measures take as long as 3 min per image (1200 pixels × 1600 pixels).(13) Our method compares favorably with fully automated methods, the most recent of which has been reported to take 0.6–0.9 s per image (1200 pixels × 1600 pixels).(13) Given algorithm design, it is unlikely that the observed differences in processing time are due in majority to image size differences.

Finally, we present a guide to how many chords to measure per field of view, given an estimation of MLI for that field. This can be used in an iterative process to tune the number of test lines used per image to best estimate MLI for that image. These numerical guidelines can be used to determine a suitably high magnification for image analysis, but, importantly, should only be applied to measures in pixel units.

There are no accepted values for the amount of lung to sample. While we present no guidelines on how much lung to sample, it would not be difficult to assess the entire lung specimen using a block-processing algorithm. According to the “do more less well” principle, it is cost ineffective to expend significant resources to ensure the genuineness of each intercept [35]. With the emphasis of our method on speed, there is effectively no cost to achieving high local accuracy while sampling a larger percentage of the lung. The paradigm of whole-slide processing gives new meaning to the “do more less well” principle. High accuracy per image/block (high local accuracy) is the only parameter of interest in a discussion of accurately measuring MLI for the whole lung. In fact, the distribution of chord lengths could be mapped across the lung, allowing a more comprehensive characterization of pathology than just a single MLI value per specimen. Our guide is meant to answer the question of how much is enough to ensure local accuracy and to control for image processing methods used.

As for commentary on usage of the guide: if MLI is small, it is probably worth using higher magnification to achieve better resolving power, as at this size of MLI, non-random effects due to thresholding methods become non-negligible, Table 1. We consistently observed this non-random effect in our set of computer-generated images and this effect is why we observe AUC_ROC < 0.5 for the 5- and 10-pixel radius strata. Also, the proposed numbers of chords are overestimates compared to what would be necessary when measuring lung morphology. This effect is due to the negative skew of the distribution of individual chord lengths in uniform-radius fields compared to the positive skew of the same distribution for random-radius circles and actual lung. In random-radius circles and lung tissue, longer chords make up a larger percentage of the total chords measured, whereas the upper bound of chord lengths in uniform-radius circle images is set by the diameter of the circle, and indicates that a higher portion of chord lengths will be in the 5- or 10-pixel range. This distributional phenomenon is reflected by the lower optimal number of chords found necessary when measuring random-radius circles, Table 2.

There are a number of limitations of the parameter MLI which have been discussed extensively elsewhere [1, 3]. An alternative proposed by Parameswaran et al. can be readily calculated from the distribution of individual chords yielded by the present method [2]. In this article, we present only a method of estimation of MLI. Importantly, we optimize the use of a previously established system of test lines to minimize time and tedium; therefore, the limitations of the test-line system are beyond the scope of this paper. For a thorough discussion of these limitations, the reader is referred to the following articles [3, 9]. Finally, in application to a murine model, we acknowledge that there are a number of fixation method parameters and lung sampling procedures that can be optimized for different purposes. These considerations are outside the scope of this paper and are discussed elsewhere [1].

Our method has several limitations and a number of application parameters are important to consider. These limitations include the image resolution—the resolving power of the method is 1 pixel; therefore, it is important to use images under sufficiently high magnification, where the 1-pixel resolving power is negligible compared to the average length of the chords to be measured.

Fully automated methods consistently underestimate MLI compared to manual measures [13, 14]. We do not know how our method compares to manual measurements in terms of under- or overestimation of MLI. Additionally, we present no method of eliminating blood vessels and other non-alveolar structures present on histology. There is arguably little worth in pursuing such structure recognition, due to its high computational cost and the “do more less well” principle as previously discussed. Because hundreds of chords can be measured per image, chord measurements produced by these structures are negligible.

In conclusion, our semi-automated MLI measurement method allows for quantification of a large area of lung, and relies on freely available software. The method produced reliable, accurate measurements and compares favorably with other methods in terms of computational requirements and speed. We successfully applied a modified version of the method to murine histology to show anatomical changes. The modification was simply in the sense of integration of multiple software platforms; the method was later streamlined to use only one platform. Given the nature of the method, it could be easily adapted to make serial measures of other types of tissue. Future applications include implementation as a premier method of histological quantification in a laboratory setting.