Automated Computational Pathology to Assess Degenerative Disc Histology

Abstract

Background

Preclinical models of disc degeneration are important tools to discover disease pathology. Histopathology is often used to understand these changes, but analyses remain reliant on pathologists or graders using time-consuming scoring systems. The integration of computational pathology can improve this process by leveraging machine learning (ML) algorithms. Thus, this work aimed to develop a segmentation model to identify seven distinct disc tissues and utilize the segmented tissue areas generated from the model, along with other derived measures, to estimate pathological changes that align with traditional histological scoring.

Methods

Hematoxylin and eosin-stained motion segment sections were collected from four independent studies. Each study included a disc injury puncture in Sprague Dawley rats. An active learning technique and a trained deep convolutional neural network were used to infer tissue segmentation. The model was then applied to untrained images to infer tissue segmentation, extract geometric and cell count features, and correlate these measurements with histologic scores from a standard scoring system.

Results

The segmentation model was highly performant with an Intersection over Union (mIOU) and frequency weighted Intersection over Union (fwIOU) of 0.83 ± 0.04 and 0.94 ± 0.02 in the Test set, respectively. The ML-derived measures correlated well with histologic scores, with absolute ranges from rho = 0.65 to 0.87. Further, these ML-derived measures were altered with disc degeneration with significant differences in NP cell number, NP area ratio, NP/AF border, NP roundness, and AF perimeter. Lastly, our model could measure additional tissue changes not captured through a standard histological scoring system.

Conclusions

Herein, we developed the first computational pathology model to phenotype disc degeneration tissue. Our model significantly correlates with traditional histopathology scoring methods, detects subtle differences between groups by directly measuring pathologic features in the images, and increases efficiency by automating the majority of the process.