Diseased skeletal muscle expresses mononuclear cell infiltration in the regions of perimysium. Accurate annotation or segmentation of perimysium can help biologists and clinicians to determine individualized patient treatment and allow for reasonable prognostication. However, manual perimysium annotation is time consuming and prone to inter-observer variations. Meanwhile, the presence of ambiguous patterns in muscle images significantly challenge many traditional automatic annotation algorithms. In this paper, we propose an automatic perimysium annotation algorithm based on deep convolutional neural network (CNN). We formulate the automatic annotation of perimysium in muscle images as a pixel-wise classification problem, and the CNN is trained to label each image pixel with raw RGB values of the patch centered at the pixel. The algorithm is applied to 82 diseased skeletal muscle images. We have achieved an average precision of 94% on the test dataset.
Morphological characteristics of muscle cells, such as cross sectional areas (CSAs), are critical factors to determine the muscle health. Automatic muscle fiber segmentation is often the first prerequisite. However, it is challenging to achieve effective and efficient skeletal muscle cell segmentation on Hematoxylin and Eosin (H&E) stained muscle images due to the complex nature of histopathology imaging and a large number of cells on a single image. In this paper, we propose to formulate the cell segmentation as a pixel-wise classification problem and train a deep convolutional neural network (CNN) to segment out the cell boundaries. Considering the speed, we apply the CNN model training to multiple graphical processing units (GPUs), and implement the distributed testing on the Spark parallel computing platform. To further improve running time cost, we apply a fast scanning technique to the pixel-wise classication with the learned CNN model. We have presented the segmentation results on a set of 120 H& E stained muscle images using the trained CNN model, and evaluated the proposed framework with accuracy calculation and speed performance.