Abstract

Lung cancer diagnosis is a critical healthcare issue, and fully automated lung nodule detection is desirable for a timely diagnosis. However, due to the variability in shapes, sizes, textures, and locations in lung nodules, developing a computer vision system for this detection is a very challenging task.
In this thesis, a special CNN system is proposed for lung nodule detection. It consists of 2 stages, namely Stage A and Stage B. Stage A is designed to localize the nodule candidates, aiming at a high sensitivity in order to minimise the miss rate. Stage B is to identify the true nodules from the input samples. It can be used to identify falsely detected nodule samples from the output of Stage A, and also as a stand-alone lung nodule recognition system.
In Stage A, there are three blocks, i.e., a pre-processing block, custom-design CNN block and refinement block. The core of this stage is the custom-designed and U-net-based CNN block. The filtering modules in its convolution layers are specifically designed to suit the features produced in these layers. To reduce the data loss in the first four layers, Full-ReLU is used as the activation function. Furthermore, the refinement block is placed to reduce effectively the false positive rate. The computation complexity of Stage A is very low, as its total number of trainable parameters is only 0.16 M. Stage A delivers a high detection rate of 95.38% but the false positive rate is still as high as 6.9 FPs/scan. The output data will be applied to Stage B for further processing.
The design of Stage B is focused on distinguishing between the true nodules and their look-likes. Based on our analysis on the characters carried by nodules of different sizes, we propose to have 2 networks in Stage B for large and small nodule categories, respectively. The feature extraction in the 2 CNNs should be different, one targeting the variations in object regions of large nodules and the other looking more into nodule surroundings in case of small nodules. Two CNNs have been designed and each of them has a particular multi-branch feature extraction (FE) block for the designated nodule category. Each CNN also involves fully-connected layers for classification. Stage B has been tested as a stand-alone lung nodule recognition system on LUNA 16 dataset. The results demonstrate that, with respect to similar systems found recently in literature, Stage B provides a good processing quality at a computation cost that is only a very small fraction of that needed by others.
The complete system for lung nodule detection, i.e., Stage A and Stage B combined, has also been tested with the same dataset. The results demonstrate the good functionality of the system. All these CNNs combined require 0.7M parameters, far less that other CNN systems performing the same task.
In summary, the proposed system has been custom-designed to optimize the computation efficiency, i.e., achieving a good detection quality at the lowest computation cost. To attain this goal, the design strategy is to decompose the complex task of lung nodule detection into subtasks so that the system can employs multiple simple CNNs, each performing a sub-task. In this way, each CNN can be structured to suit the characters of a particular kind of nodule data and optimised to meet specific performance requirements. The effectiveness of this strategy has been confirmed by the results of the performance evaluation. Because of its low computation cost, the proposed system can be very easily implemented in various environment.