Towards an affordable deep learning system: automated intervertebral disc detection in x-ray images

Adult Spinal Deformity is a prominent medical issue with about 68% of the elderly population suffering from the disease.(1) Detailed biomechanical assessment is needed both in the presurgical planning of structural spinal deformity as well as in early functional biomechanical compensation in ambulatory spinal pain patients. When considering automation of this process, we have to look at photographic intervertebral disc detection technique as a way to produce a detailed model of the spine with appropriate measurements required to make efficient and accurate decisions on patient care. Deep convolutional neural network (CNN) has given remarkable results in object recognition tasks in recent years. However, massive training data, computational resources and long training time is needed for both training a deep network from scratch or finetuning a network. Using pretrained model as feature extractor has shown promising result for moderate sized medical data.2 However, most work have extracted features from the last layer and little has been explored in terms of the number of convolutional layers needed for best performance. In this work we trained Support Vector Machine (SVM) classifiers on different layers of CaffeNet3 features to show that deeper the better concept does not hold for task such as intervertebral disc detection. Furthermore, our experimental results show the potential of using very small training data, such as 15 annotated medical images in our experiment, to yield satisfactory classification performance with accuracy up to 97.2%