Automated/Quantitative assessment of anatomical specimen preservation techniques using machine learning.

Biological tissue preservation has historically been an area of interest for scientific purposes. Concerning medical education broad options are available thanks to preservation techniques. Soft tissue plastination is a preservation technique capable of providing dry, odorless and durable specimens. It has multiple variations, each one with particular characteristics. Nevertheless, comparing each preservation method is usually subjective to the observer's opinion and there are no published automated quantitative tools for this assessment. In the last decade, machine learning techniques have shown promising performance in image processing and computer vision problems. Applying these algorithms to evaluate multiple preservation methods can change the way researchers judge their work. Our aim is to provide a framework for educators and researchers for quantitatively measuring performance and similarity of preparation techniques, given a single image of the result in terms of color and tissue differentiation. As a first approach, 68 porcine heart axial‐cut images from four categories were acquired: fresh (F), resin sheet plastinated, silicone plastinated and low Formaldehyde (LF) preserved. F and LF specimens' images were used as controls for best and worst color and tissue differentiation examples respectively. Muscle/fat tissue ground truth was pursued for each category by a last year medical student. All data were pre‐processed to improve processing time and to provide normalized input for the algorithms. Support vector machines (SVM) were used to assess performance by solving a pixel classification task (segmentation of fat and muscular tissue). As features of the images we used the CIELAB color space to represent each pixel, hence the dimensionality of the input instances (LAB) is d=3. In this representation space, we used a Gaussian kernel for non‐linear input instances mapping into higher dimensional features, this approach has shown to dramatically improve the training algorithm performance. Fold cross validation was used to select SVM and kernel optimal parameters, thus improving confidence and average precision, additionally we used F‐score to find the best threshold for tissue segmentation. Inner product was used as a measure of resemblance between each method. Representation used for each image were the histograms in each CIELAB channels. Thus, input dimension instances are the concatenation of the histogram (with 128 bins) for each class d= 384. Couple permutations without repetition were taken and average over each equivalent to obtain a better estimate of the similarity between each plastination method and the controls. Predictions for all methods were highly successful, showing average precisions between 95% and 98%. Hence the algorithm performance was no only good but also improves the state of the art in user evaluation, this way educators and researchers can measure specimens' quality. In terms of similarity, both plastination methods are closer to F than LF preserved hearts. However, between both methods silicone based plastination is closer to F. The final SVM model is capable of quantitatively analyze any related specimen images, by comparing the input image with other examples of images of the same method. It can also asses the distance between the input image and the positive and negative controls. This abstract is from the Experimental Biology 2018 Meeting. There is no full text article associated with this abstract published in The FASEB Journal .