As the title suggests, this paper studies various different deep convolutional neural network architectures and various techniques to use these CNNs for CADe (Computer Aided Detection) tasks. With the tremendous popularity of CNN models, the authors state that the “tremendous” success of CNNs in medical image tasks has been primarily using three techniques.
- Training the CNN from “scratch”.
- Using “off-the-shelf” CNN features to use alongside hand crafted image features.
- Using pre-trained (on natural images) neural networks and fine-tuning them to medical images using transfer learning.
Two CADe problems, namely thoraco-abdominal lymph node (LN) detection and interstitial lung disease (ILD) classification have been addressed in the paper, and three popular models (along with their variants) - CifarNet (shallow network), AlexNet (7-layered network), and GoogLeNet (very deep network) - have been studied. Moreover, the authors study the effect of different varying dataset scale and spatial image context on performance, while evaluating networks with a wide range of number of parameters (5 thousand to 160 million), and also explore when and why transfer learning from pre-trained ImageNet CNN models might prove useful.
The LN dataset consists of 388 mediastinal LNs labeled by radiologists in 90 patient CT scans, and 595 abdominal LNs in 86 patient CT scans. The ILD dataset is a publicly available one, which comprises of 905 image slices from 120 patients, with six lung tissue types.
CifarNet has been used as a baseline and AlexNet and GoogLeNet and their modifications are evaluated. In particular, high resolution and low resolution versions of the two networks have been evaluated. The authors hypothesize that despite the dissimilarities between natural and artificial images, the CNNs trained ‘comprehensively’ on a large well annotated natural images’ dataset such as ImageNet may still be transferred to perform efficiently on medical image related tasks.
The authors summarize their observations of the various results as the following conclusions:
When the size of the training dataset is limited, deep CNN architectures are useful for CADe tasks. This is especially important because the contemporary CNNs being used for medical image tasks.
There is a trade-off between using better learning models versus more training data, and the optimal trade-off varies task-to-task.
Limited datasets can prove to be a bottleneck for CADe tasks, and building progressively larger and well annotated datasets are equally, if not more, important than designing better algorithms.
Transfer learning is an efficient approach for using information from large well annotated datasets (such as ImageNet) for CADe problems.
“Off-the-shelf” features from deep CNN models can be applied to CADe applications in a better way by exploring “performance-complementary properties of hand-crafted features” or by fine-tuning the pre-trained models on the target medical image dataset.